Models of vegetation biomass based on climate predictors

Based on code from Martin Holdrege’s Sagebrush-Fire paper

The data consists of vegetation % cover by functional group from across CONUS (from AIM, FIA, LANDFIRE, and RAP), as well as climate variables from DayMet, which have been aggregated into mean interannual conditions accross multiple temporal windows.

Dependencies

Set user defined parameters

print(params)

## $run
## [1] FALSE
## 
## $save_figs
## [1] FALSE
## 
## $ecoregion
## [1] "CONUS"
## 
## $response
## [1] "RAP"
## 
## $removeTexasLouisianaPlain
## [1] FALSE
## 
## $trimAnomalies
## [1] TRUE
## 
## $autoKfold
## [1] FALSE

# set to true if want to run for a limited number of rows (i.e. for code testing)
test_run <- params$test_run
save_figs <- params$save_figs
response <- params$response
fit_sample <- TRUE # fit model to a sample of the data
n_train <- 5e4 # sample size of the training data
n_test <- 1e6 # sample size of the testing data (if this is too big the decile dotplot code throws memory errors)
trimAnom <- params$trimAnomalies
removeTLP <- params$removeTexasLouisianaPlain
run <- params$run
autoKfold <- params$autoKfold

Load packages

# set option so resampled dataset created here reproduces earlier runs of this code with dplyr 1.0.10
source("../../../Functions/glmTransformsIterates.R")
source("../../../Functions/transformPreds.R")
source("../../../Functions/StepBeta_mine.R")
#source("src/fig_params.R")
#source("src/modeling_functions.R")
 
library(ggspatial)
library(terra)
library(tidyterra)
library(sf)
library(caret)
library(tidyverse)
library(GGally) # for ggpairs()
library(pdp) # for partial dependence plots
library(gridExtra)
library(knitr)
library(patchwork) # for figure insets etc. 
library(ggtext)
library(StepBeta)
theme_set(theme_classic())
library(here)
library(rsample)
library(kableExtra)
library(glmnet)
library(USA.state.boundaries)

Read in data

Data compiled in the prepDataForModels.R script

here::i_am("Analysis/VegComposition/ModelFitting/02_ModelFitting.Rmd")
modDat <- readRDS("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Data_processed/BiomassQuantityData/dataForAnalysis_fireAndDevelopmentRemoved.rds")

modDat_1 <- modDat %>% 
  st_drop_geometry()
# remove second 'x' and 'y' columns
#modDat <- modDat[,c(1:41,43:48)]
# For all response variables, make sure there are no 0s add  .0001 from each, since the Gamma model framework can't handle that
modDat_1[modDat_1$biomass_MgPerHect == 0 & !is.na(modDat_1$biomass_MgPerHect), "biomass_MgPerHect"] <- 0.0001

Prep data

Add a constant to the response variable (+2) so that models run…

modDat_1 <- modDat_1 %>%
mutate(response_transformed = .[["biomass_MgPerHect"]] + 2)

Identify the ecoregion and response variable type to use in this model run

ecoregion <- params$ecoregion
response <- params$response
print(paste0("In this model run, the ecoregion is ", ecoregion," and the data is from ",response))

## [1] "In this model run, the ecoregion is CONUS and the data is from RAP"

Visualize the predictor variables

The following are the candidate predictor variables for this ecoregion:

if (ecoregion == "shrubGrass") {
  # select potential predictor variables for the ecoregion of interest
        prednames <-
          c("tmean", "prcpTempCorr", "isothermality", "annWatDef",
         "prcp", "prcp_seasonality", "prcp_dry", "annWetDegDays",
         "t_warm", "t_cold", "prcp_wet", "soilDepth", "sand", "coarse",
         "AWHC", "clay", "carbon"
# "tmean"             , "prcp"                    ,"prcp_seasonality"        ,"prcpTempCorr"          , 
# "isothermality"     , "annWatDef"               ,"sand"                    ,"coarse"                , 
# "carbon"            , "AWHC"                    ,"tmin_anom"               ,"tmax_anom"             , 
# "t_warm_anom"       , "prcp_wet_anom"           ,"precp_dry_anom"          ,"prcp_seasonality_anom" , 
# "prcpTempCorr_anom" , "aboveFreezingMonth_anom" ,"isothermality_anom"      ,"annWatDef_anom"        , 
# "annWetDegDays_anom", "VPD_mean_anom"           ,"VPD_min_anom"            ,"frostFreeDays_5_anom"  
)
  
} else if (ecoregion %in% c("forest", "eastForest", "westForest")) {
  # select potential predictor variables for the ecoregion of interest
  prednames <- 
    c("tmean", "prcpTempCorr", "isothermality", "annWatDef",
         "prcp", "prcp_seasonality", "prcp_dry", "annWetDegDays",
         "t_warm", "t_cold", "prcp_wet", "soilDepth", "sand", "coarse",
         "AWHC", "clay", "carbon"
# "tmean"                 ,"prcp"               , "prcp_dry"                , "prcpTempCorr"      ,     
# "isothermality"         ,"annWatDef"          , "clay"                    , "sand"              ,     
# "coarse"                ,"carbon"             , "AWHC"                    , "tmin_anom"         ,     
# "tmax_anom"             ,"prcp_anom"          , "prcp_wet_anom"           , "precp_dry_anom"    ,     
# "prcp_seasonality_anom" ,"prcpTempCorr_anom"  , "aboveFreezingMonth_anom" , "isothermality_anom",     
# "annWatDef_anom"        ,"annWetDegDays_anom" , "VPD_mean_anom"           , "VPD_max_95_anom"   ,     
# "frostFreeDays_5_anom"   
)
} else if (ecoregion == "CONUS") {
  prednames <- c("tmean", "prcpTempCorr", "isothermality", "annWatDef",
         "prcp", "prcp_seasonality", "prcp_dry", "annWetDegDays",
         "t_warm", "t_cold", "prcp_wet", "soilDepth", "sand", "coarse",
         "AWHC", "clay", "carbon"
#     "tmean"               ,"prcp"               ,"prcp_seasonality", "prcpTempCorr"       ,  "isothermality"     ,     
# "annWetDegDays"           ,"sand"               ,"coarse"         , "AWHC"                , "tmin_anom"          ,    
# "tmax_anom"               ,"prcp_wet_anom"      ,"precp_dry_anom" , "prcp_seasonality_anom", "prcpTempCorr_anom" ,     
# "aboveFreezingMonth_anom" ,"isothermality_anom" ,"annWatDef_anom" , "annWetDegDays_anom"  , "VPD_mean_anom"      ,    
# "VPD_max_95_anom"         ,"frostFreeDays_5_anom"   
  )
} 

# get the name of the transformed response
#response <- paste0(response, "_transformed")

Scale the predictor variables for the model-fitting process

allPreds <- modDat_1 %>% 
  dplyr::select(tmin:frostFreeDays,#tmean_anom:frostFreeDays_anom, 
                soilDepth:AWHC) %>% 
  names()
modDat_1_s <- modDat_1 %>% 
  mutate(across(all_of(allPreds), base::scale, .names = "{.col}_s")) 
# names(modDat_1_s) <- c(names(modDat_1),
#                        paste0(prednames, "_s")
#                        )
#save model input data after its been scaled
saveRDS(modDat_1_s, file = "./models/scaledModelInputData.rds")

Isolate the data just for the response variable of interest

modDat_1 <- modDat_1 %>% 
  filter(biomassSource == response)
modDat_1_s <- modDat_1_s %>% 
  filter(biomassSource == response)

Subset the data to only include data for the ecoregion of interest

if (ecoregion == "shrubGrass") {
  # select data for the ecoregion of interest
  modDat_1_s <- modDat_1_s %>%
    filter(newRegion == "dryShrubGrass")
} else if (ecoregion == "forest") {
  # select data for the ecoregion of interest
  modDat_1_s <- modDat_1_s %>% 
    filter(newRegion %in% c("eastForest", "westForest"))
} else if (ecoregion == "CONUS") {
  modDat_1_s <- modDat_1_s
} else if (ecoregion == "eastForest") {
   modDat_1_s <- modDat_1_s %>% 
    filter(newRegion == "eastForest")
} else if (ecoregion == "westForest") {
    modDat_1_s <- modDat_1_s %>% 
    filter(newRegion == "westForest")
}
# remove the rows that have no observations for the response variable of interest
modDat_1_s <- modDat_1_s[!is.na(modDat_1_s$biomass_MgPerHect),]

# subset the data to only include these predictors, and remove any remaining NAs 
modDat_1_s <- modDat_1_s %>% 
  dplyr::select(prednames, paste0(prednames, "_s"), biomass_MgPerHect, response_transformed, 
                newRegion, #Year, Long, Lat,
                NA_L1NAME, NA_L2NAME,
                Long, Lat) %>% 
  drop_na()

names(prednames) <- prednames
df_pred <- modDat_1_s[, prednames]
# 
# # print the list of predictor variables
# knitr::kable(format = "html", data.frame("Possible_Predictors" = prednames), row.names = FALSE
# ) %>%
#   kable_styling(bootstrap_options = c("striped", "hover", "condensed"))

Visualize the response variable

hist(modDat_1_s$biomass_MgPerHect, main = paste0("Histogram of ","biomass_MgPerHect"),
     xlab = paste0("biomass_MgPerHect"))

create_summary <- function(df) {
  df %>% 
    pivot_longer(cols = everything(),
                 names_to = 'variable') %>% 
    group_by(variable) %>% 
    summarise(across(value, .fns = list(mean = ~mean(.x, na.rm = TRUE), min = ~min(.x, na.rm = TRUE), 
                                        median = ~median(.x, na.rm = TRUE), max = ~max(.x, na.rm = TRUE)))) %>% 
    mutate(across(where(is.numeric), round, 4))
}

modDat_1_s[prednames] %>% 
  create_summary() %>% 
  knitr::kable(caption = 'summaries of possible predictor variables') %>%
kable_styling(bootstrap_options = c("striped", "hover", "condensed")) 

summaries of possible predictor variables

variable	value_mean	value_min	value_median	value_max
AWHC	17.7939	0.1200	17.5384	36.8252
annWatDef	103.8234	0.5966	72.3612	551.1612
annWetDegDays	2206.2726	111.7359	2004.6685	6767.0468
carbon	1.2267	0.0000	0.9581	34.6588
clay	21.4685	1.4289	20.8353	82.0619
coarse	6.1422	0.0000	2.6768	82.9684
isothermality	36.0466	21.6565	36.1507	60.1487
prcp	519.1832	48.5513	452.3900	1749.9463
prcpTempCorr	0.1585	-0.7966	0.2784	0.6947
prcp_dry	3.7026	0.0000	2.2250	37.2280
prcp_seasonality	0.9705	0.5220	0.9300	2.2437
prcp_wet	127.4978	22.3970	116.3538	500.3483
sand	42.1250	0.3922	40.6378	99.0941
soilDepth	151.7315	3.0000	161.9290	201.0000
t_cold	-7.8513	-22.9577	-8.5266	11.0941
t_warm	32.5296	10.3896	32.1547	47.3658
tmean	11.9035	-4.8245	10.8338	24.8209

# response_summary <- modDat_1_s %>% 
#     dplyr::select(#where(is.numeric), -all_of(pred_vars),
#       matches(response)) %>% 
#     create_summary()
# 
# 
# kable(response_summary, 
#       caption = 'summaries of response variables, calculated using paint') %>%
# kable_styling(bootstrap_options = c("striped", "hover", "condensed")) 

Histograms of raw and scaled predictors

scaleFigDat_1 <- modDat_1_s %>% 
  dplyr::select(c(#Long, Lat, Year, 
    Long, Lat, prednames)) %>% 
  pivot_longer(cols = all_of(names(prednames)), 
               names_to = "predNames", 
               values_to = "predValues_unScaled")
scaleFigDat_2 <- modDat_1_s %>% 
  dplyr::select(c(#Long, Lat, Year, 
    Long, Lat, paste0(prednames, "_s"))) %>% 
  pivot_longer(cols = all_of(paste0(prednames,"_s"
                                    )), 
               names_to = "predNames", 
               values_to = "predValues_scaled", 
               names_sep = ) %>% 
  mutate(predNames = str_replace(predNames, pattern = "_s$", replacement = ""))

scaleFigDat_3 <- scaleFigDat_1 %>% 
  left_join(scaleFigDat_2)

ggplot(scaleFigDat_3) + 
  facet_wrap(~predNames, scales = "free") +
  geom_histogram(aes(predValues_unScaled), fill = "lightgrey", col = "darkgrey") + 
  geom_histogram(aes(predValues_scaled), fill = "lightblue", col = "blue") +
  xlab ("predictor variable values") + 
  ggtitle("Comparing the distribution of unscaled (grey) to scaled (blue) predictor variables")

modDat_1_s <- modDat_1_s %>% 
  rename_with(~paste0(.x, "_raw"), 
                any_of(names(prednames))) %>% 
  rename_with(~str_remove(.x, "_s$"), 
              any_of(paste0(names(prednames), "_s")))

Predictor variables compared to binned response variables

set.seed(12011993)
quants <- quantile(modDat_1_s$biomass_MgPerHect)
# vector of name of response variables
vars_response <- "biomass_MgPerHect"
# longformat dataframes for making boxplots
df_sample_plots <-  modDat_1_s  %>% 
  slice_sample(n = 5e4) %>% 
   #rename("biomass_MgPerHect" = all_of(biomass_MgPerHect)) %>% 
  mutate("biomass_MgPerHect" = case_when(
    "biomass_MgPerHect" <= quants[2] ~ paste(round(quants[2],2)), 
    "biomass_MgPerHect" > quants[2] & biomass_MgPerHect <=quants[3] ~ paste(round(quants[3],2)), 
    "biomass_MgPerHect" > quants[3] & biomass_MgPerHect <=quants[4] ~ paste(round(quants[4],2)), 
    "biomass_MgPerHect" >= quants[4]  ~ paste(round(quants[5],2)), 
  )) %>% 
  dplyr::select(c("biomass_MgPerHect", prednames)) %>% 
  tidyr::pivot_longer(cols = unname(prednames), 
               names_to = "predictor", 
               values_to = "value"
               )  
 

  ggplot(df_sample_plots, aes_string(x= "biomass_MgPerHect", y = 'value')) +
  geom_boxplot() +
  facet_wrap(~predictor , scales = 'free_y') + 
  ylab("Predictor Variable Values") + 
    xlab("biomass_MgPerHect")

Model Fitting

Visualize the spatial blocks and how they differ across environmental space

First, if there are observations in Louisiana, sub-sample them so they’re not so over-represented in the dataset

## make data into spatial format
modDat_1_sf <- modDat_1_s %>% 
  st_as_sf(coords = c("Long", "Lat"), crs = st_crs("PROJCRS[\"unnamed\",\n    BASEGEOGCRS[\"unknown\",\n        DATUM[\"unknown\",\n            ELLIPSOID[\"Spheroid\",6378137,298.257223563,\n                LENGTHUNIT[\"metre\",1,\n                    ID[\"EPSG\",9001]]]],\n        PRIMEM[\"Greenwich\",0,\n            ANGLEUNIT[\"degree\",0.0174532925199433,\n                ID[\"EPSG\",9122]]]],\n    CONVERSION[\"Lambert Conic Conformal (2SP)\",\n        METHOD[\"Lambert Conic Conformal (2SP)\",\n            ID[\"EPSG\",9802]],\n        PARAMETER[\"Latitude of false origin\",42.5,\n            ANGLEUNIT[\"degree\",0.0174532925199433],\n            ID[\"EPSG\",8821]],\n        PARAMETER[\"Longitude of false origin\",-100,\n            ANGLEUNIT[\"degree\",0.0174532925199433],\n            ID[\"EPSG\",8822]],\n        PARAMETER[\"Latitude of 1st standard parallel\",25,\n            ANGLEUNIT[\"degree\",0.0174532925199433],\n            ID[\"EPSG\",8823]],\n        PARAMETER[\"Latitude of 2nd standard parallel\",60,\n            ANGLEUNIT[\"degree\",0.0174532925199433],\n            ID[\"EPSG\",8824]],\n        PARAMETER[\"Easting at false origin\",0,\n            LENGTHUNIT[\"metre\",1],\n            ID[\"EPSG\",8826]],\n        PARAMETER[\"Northing at false origin\",0,\n            LENGTHUNIT[\"metre\",1],\n            ID[\"EPSG\",8827]]],\n    CS[Cartesian,2],\n        AXIS[\"easting\",east,\n            ORDER[1],\n            LENGTHUNIT[\"metre\",1,\n                ID[\"EPSG\",9001]]],\n        AXIS[\"northing\",north,\n            ORDER[2],\n            LENGTHUNIT[\"metre\",1,\n                ID[\"EPSG\",9001]]]]"))


# download map info for visualization
data(state_boundaries_wgs84) 

cropped_states <- suppressMessages(state_boundaries_wgs84 %>%
  dplyr::filter(NAME!="Hawaii") %>%
  dplyr::filter(NAME!="Alaska") %>%
  dplyr::filter(NAME!="Puerto Rico") %>%
  dplyr::filter(NAME!="American Samoa") %>%
  dplyr::filter(NAME!="Guam") %>%
  dplyr::filter(NAME!="Commonwealth of the Northern Mariana Islands") %>%
  dplyr::filter(NAME!="United States Virgin Islands") %>%
  sf::st_sf() %>%
  sf::st_transform(sf::st_crs(modDat_1_sf))) #%>%
  #sf::st_crop(sf::st_bbox(modDat_1_sf)+c(-1,-1,1,1))
if (ecoregion %in% c("Forest", "eastForest", "forest")){
modDat_1_s$uniqueID <- 1:nrow(modDat_1_s)
modDat_1_sf$uniqueID <- 1:nrow(modDat_1_sf)

  #xlim(c(730439.1, 1042196))
}

## do a pca of climate across level 2 ecoregions
pca <- prcomp(modDat_1_s[,paste0(prednames)])
library(factoextra)
(fviz_pca_ind(pca, habillage = modDat_1_s$NA_L2NAME, label = "none", addEllipses = TRUE, ellipse.level = .95, ggtheme = theme_minimal(), alpha.ind = .1))

if (ecoregion == "shrubGrass") {
  print("We'll combine the 'Mediterranean California' and 'Western Sierra Madre Piedmont' ecoregions (into 'Mediterranean Piedmont'). We'll also combine `Tamaulipas-Texas semiarid plain,' 'Texas-Lousiana Coastal plain,' and 'South Central semiarid prairies' ecoregions (into (`Semiarid plain and prairies`)." )
  
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MEDITERRANEAN CALIFORNIA", "WESTERN SIERRA MADRE PIEDMONT"), "NA_L2NAME"] <- "MEDITERRANEAN PIEDMONT"
  
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("TAMAULIPAS-TEXAS SEMIARID PLAIN", "TEXAS-LOUISIANA COASTAL PLAIN", "SOUTH CENTRAL SEMIARID PRAIRIES"), "NA_L2NAME"] <- "SEMIARID PLAIN AND PRAIRIES"
 
    modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND WEST COAST FOREST"
  
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MEDITERRANEAN CALIFORNIA", "UPPER GILA MOUNTAINS"), "NA_L2NAME"] <- "MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS"

modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD PLAINS", "OZARK/OUACHITA-APPALACHIAN FORESTS"), "NA_L2NAME"] <- "OZARK/OUACHITA-APPALACHIAN FORESTS AND MIXED WOOD PLAINS"
#////
modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MEDITERRANEAN CALIFORNIA", "WESTERN SIERRA MADRE PIEDMONT"), "NA_L2NAME"] <- "MEDITERRANEAN PIEDMONT"
  
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("TAMAULIPAS-TEXAS SEMIARID PLAIN", "TEXAS-LOUISIANA COASTAL PLAIN", "SOUTH CENTRAL SEMIARID PRAIRIES"), "NA_L2NAME"] <- "SEMIARID PLAIN AND PRAIRIES"
 
    modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND WEST COAST FOREST"
  
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MEDITERRANEAN CALIFORNIA", "UPPER GILA MOUNTAINS"), "NA_L2NAME"] <- "MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS"

modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD PLAINS", "OZARK/OUACHITA-APPALACHIAN FORESTS"), "NA_L2NAME"] <- "OZARK/OUACHITA-APPALACHIAN FORESTS AND MIXED WOOD PLAINS"
 if (response == "CAMCover") {
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MEDITERRANEAN PIEDMONT", "SEMIARID PLAIN AND PRAIRIES"), "NA_L2NAME"] <- "SEMIARID PLAIN AND PRAIRIES AND PIEDMONT"
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MEDITERRANEAN PIEDMONT", "SEMIARID PLAIN AND PRAIRIES"), "NA_L2NAME"] <- "SEMIARID PLAIN AND PRAIRIES AND PIEDMONT"
 } else if (response %in% c("C4GramCover_prop", "C3GramCover_prop")) {
     modDat_1_s[modDat_1_s$NA_L2NAME %in% c("SEMIARID PLAIN AND PRAIRIES", "TEMPERATE PRAIRIES"), "NA_L2NAME"] <- "SEMIARID AND TEMPERATE PRAIRIES"
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("SEMIARID PLAIN AND PRAIRIES", "TEMPERATE PRAIRIES"), "NA_L2NAME"] <- "SEMIARID AND TEMPERATE PRAIRIES" 
 }

} else if (ecoregion == "CONUS") {
  
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("EVERGLADES", "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"), "NA_L2NAME"] <- "EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"
  
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND WEST COAST FOREST"
  
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD PLAINS", "OZARK/OUACHITA-APPALACHIAN FORESTS"), "NA_L2NAME"] <- "OZARK/OUACHITA-APPALACHIAN FORESTS AND MIXED WOOD PLAINS"
  
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MEDITERRANEAN CALIFORNIA", "UPPER GILA MOUNTAINS"), "NA_L2NAME"] <- "MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS"
  
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("OZARK/OUACHITA-APPALACHIAN FORESTS AND MIXED WOOD PLAINS", "SOUTHEASTERN USA PLAINS",  "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS", "EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN AND MIXED WOOD PLAINS"
   
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("SOUTH CENTRAL SEMIARID PRAIRIES", "TEXAS-LOUISIANA COASTAL PLAIN"), "NA_L2NAME"] <- "SOUTH CENTRAL SEMIARID PRAIRIES"
  #///
  
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("EVERGLADES", "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"), "NA_L2NAME"] <- "EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"
  
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND WEST COAST FOREST"
  
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD PLAINS", "OZARK/OUACHITA-APPALACHIAN FORESTS"), "NA_L2NAME"] <- "OZARK/OUACHITA-APPALACHIAN FORESTS AND MIXED WOOD PLAINS"
  
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MEDITERRANEAN CALIFORNIA", "UPPER GILA MOUNTAINS"), "NA_L2NAME"] <- "MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS"
  
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("OZARK/OUACHITA-APPALACHIAN FORESTS AND MIXED WOOD PLAINS", "SOUTHEASTERN USA PLAINS",  "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS", "EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN AND MIXED WOOD PLAINS"
   
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("SOUTH CENTRAL SEMIARID PRAIRIES", "TEXAS-LOUISIANA COASTAL PLAIN"), "NA_L2NAME"] <- "SOUTH CENTRAL SEMIARID PRAIRIES"
  
  if (response %in% c("C4GramCover_prop")) {
    modDat_1_s[modDat_1_s$NA_L2NAME %in% c("CENTRAL USA PLAINS", "TEMPERATE PRAIRIES", "SOUTHEASTERN AND MIXED WOOD PLAINS", "ATLANTIC HIGHLANDS", "MIXED WOOD SHIELD"), "NA_L2NAME"] <- "EASTERN AND MIXED WOOD PLAINS AND FOREST"
  #///
  
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in%c("CENTRAL USA PLAINS", "TEMPERATE PRAIRIES", "SOUTHEASTERN AND MIXED WOOD PLAINS", "ATLANTIC HIGHLANDS", "MIXED WOOD SHIELD"), "NA_L2NAME"] <- "EASTERN AND MIXED WOOD PLAINS AND FOREST"
  }
} else if (ecoregion == "forest"  & response != "CAMCover") {
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS", "EVERGLADES"), "NA_L2NAME"] <- "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS    AND EVERGLADES"
   
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("UPPER GILA MOUNTAINS", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS"
   
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS\tAND EVERGLADES", "SOUTHEASTERN USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN USA PLAINS"
   
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("ATLANTIC HIGHLANDS", "OZARK/OUACHITA-APPALACHIAN FORESTS"), "NA_L2NAME"] <- "HIGHLANDS AND APPALACHIAN FORESTS"
   
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("CENTRAL USA PLAINS", "MIXED WOOD PLAINS"), "NA_L2NAME"] <- "CENTRAL AND MIXED WOOD PLAINS"
   
   modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD SHIELD", "CENTRAL AND MIXED WOOD PLAINS"), "NA_L2NAME"] <- "CENTRAL AND MIXED WOOD PLAINS AND MIXED WOOD SHIELD"
   
       ## divide southeastern US plains into two regions, since it's by far the largest group
  modDat_1_s[modDat_1_s$NA_L2NAME == "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS" &
               modDat_1_s$Long < -966595#-1773969
               , "NA_L2NAME"] <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 1"
  modDat_1_s[modDat_1_s$NA_L2NAME == "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS", "NA_L2NAME"] <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 2"
   #///
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS", "EVERGLADES"), "NA_L2NAME"] <- "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS  AND EVERGLADES"
   
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("UPPER GILA MOUNTAINS", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS"
   
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS\tAND EVERGLADES", "SOUTHEASTERN USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN USA PLAINS"
   
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("ATLANTIC HIGHLANDS", "OZARK/OUACHITA-APPALACHIAN FORESTS"), "NA_L2NAME"] <- "HIGHLANDS AND APPALACHIAN FORESTS"
   
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("CENTRAL USA PLAINS", "MIXED WOOD PLAINS"), "NA_L2NAME"] <- "CENTRAL AND MIXED WOOD PLAINS"
   
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD SHIELD", "CENTRAL AND MIXED WOOD PLAINS"), "NA_L2NAME"] <- "CENTRAL AND MIXED WOOD PLAINS AND MIXED WOOD SHIELD"
          ## divide southeastern US plains into two regions, since it's by far the largest group
  modDat_1_sf[modDat_1_sf$NA_L2NAME == "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS" &
              st_coordinates(modDat_1_sf)[,1] < -966595#-1773969
                , ]$NA_L2NAME <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 1"
  modDat_1_sf[modDat_1_sf$NA_L2NAME == "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS", "NA_L2NAME"] <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 2"
  
  if (response %in% c("C3GramCover_prop", "C4GramCover_prop") ) {
     modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 1"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 1"
     modDat_1_s[modDat_1_s$NA_L2NAME %in% c("CENTRAL AND MIXED WOOD PLAINS AND MIXED WOOD SHIELD", "HIGHLANDS AND APPALACHIAN FORESTS"), "NA_L2NAME"] <- "CENTRAL AND MIXED WOODS AND HIGHLANDS FORESTS"
     #//
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 1"), "NA_L2NAME"] <- "WESTERN CORDILLERA AND UPPER GILA MOUNTAINS 1"
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("CENTRAL AND MIXED WOOD PLAINS AND MIXED WOOD SHIELD", "HIGHLANDS AND APPALACHIAN FORESTS"), "NA_L2NAME"] <- "CENTRAL AND MIXED WOODS AND HIGHLANDS FORESTS" 
  
  }
  
} else if (ecoregion == "forest" & response == "CAMCover") {
  
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("OZARK/OUACHITA-APPALACHIAN FORESTS", "SOUTHEASTERN USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN PLAINS AND APPALACHIAN FORESTS"
    
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "MARINE WEST COAST AND WESTERN CORDILLERA"
    
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD PLAINS", "SOUTHEASTERN PLAINS AND APPALACHIAN FORESTS"), "NA_L2NAME"] <- "SOUTHEASTERN PLAINS AND APPALACHIAN FORESTS"
  
  #///
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("OZARK/OUACHITA-APPALACHIAN FORESTS", "SOUTHEASTERN USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN PLAINS AND APPALACHIAN FORESTS"
    
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MARINE WEST COAST FOREST", "WESTERN CORDILLERA"), "NA_L2NAME"] <- "MARINE WEST COAST AND WESTERN CORDILLERA"
    
  modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD PLAINS", "SOUTHEASTERN PLAINS AND APPALACHIAN FORESTS"), "NA_L2NAME"] <- "SOUTHEASTERN PLAINS AND APPALACHIAN FORESTS"
  
} else if (ecoregion == "eastForest") {
  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("EVERGLADES", "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"), "NA_L2NAME"] <- "EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"

  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD PLAINS","CENTRAL USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN AND CENTRAL USA PLAINS"

  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD SHIELD", "ATLANTIC HIGHLANDS"), "NA_L2NAME"] <- "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD"

  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD PLAINS", "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD"), "NA_L2NAME"] <- "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD AND PLAINS"

  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("SOUTHEASTERN AND CENTRAL USA PLAINS", "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD AND PLAINS"), "NA_L2NAME"] <- "PLAINS AND HIGHLANDS AND SHIELD"

  modDat_1_s[modDat_1_s$NA_L2NAME %in% c("EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS", "SOUTHEASTERN USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN PLAINS AND COAST"

  # #////
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("EVERGLADES", "MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"), "NA_L2NAME"] <- "EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS"

   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD PLAINS","CENTRAL USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN AND CENTRAL USA PLAINS"
  
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD SHIELD", "ATLANTIC HIGHLANDS"), "NA_L2NAME"] <- "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD"
  
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD PLAINS", "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD"), "NA_L2NAME"] <- "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD AND PLAINS"
  
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("SOUTHEASTERN AND CENTRAL USA PLAINS", "ATLANTIC HIGHLANDS AND MIXED WOOD SHIELD AND PLAINS"), "NA_L2NAME"] <- "PLAINS AND HIGHLANDS AND SHIELD"
  
   modDat_1_sf[modDat_1_sf$NA_L2NAME %in% c("EVERGLADES MISSISSIPPI ALLUVIAL AND SOUTHEAST USA COASTAL PLAINS", "SOUTHEASTERN USA PLAINS"), "NA_L2NAME"] <- "SOUTHEASTERN PLAINS AND COAST"
  
   ## divide southeastern US plains into two regions, since it's by far the largest group
  # modDat_1_s[modDat_1_s$NA_L2NAME == "SOUTHEASTERN USA PLAINS" &
  #              modDat_1_s$Lat < -590062, "NA_L2NAME"] <- "SOUTHEASTERN USA PLAINS 1"
  # modDat_1_s[modDat_1_s$NA_L2NAME == "SOUTHEASTERN USA PLAINS", #&
  #             # modDat_1_s$Lat < -590062,
  #            "NA_L2NAME"] <- "SOUTHEASTERN USA PLAINS 2"

  #   ## divide southeastern US plains into two regions, since it's by far the largest group
  # modDat_1_s[modDat_1_s$NA_L2NAME == "OZARK/OUACHITA-APPALACHIAN FORESTS" &
  #              modDat_1_s$Long < 854862.2, "NA_L2NAME"] <- "OZARK/OUACHITA-APPALACHIAN FORESTS 1"
  # modDat_1_s[modDat_1_s$NA_L2NAME == "OZARK/OUACHITA-APPALACHIAN FORESTS" &
  #              modDat_1_s$Long  >=   854862.2, "NA_L2NAME"] <- "OZARK/OUACHITA-APPALACHIAN FORESTS 2"
}
# make a table of n for each region
modDat_1_s %>% 
  group_by(NA_L2NAME) %>% 
  dplyr::summarize("Number_Of_Observations" = length(NA_L2NAME)) %>% 
  rename("Level_2_Ecoregion" = NA_L2NAME)%>% 
  kable() %>%
kable_styling(bootstrap_options = c("striped", "hover", "condensed")) 

Level_2_Ecoregion	Number_Of_Observations
COLD DESERTS	92503
MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS	13958
MIXED WOOD SHIELD	14
SOUTH CENTRAL SEMIARID PRAIRIES	104615
SOUTHEASTERN AND MIXED WOOD PLAINS	546
TAMAULIPAS-TEXAS SEMIARID PLAIN	5878
TEMPERATE PRAIRIES	52791
WARM DESERTS	40137
WEST-CENTRAL SEMIARID PRAIRIES	58382
WESTERN CORDILLERA AND WEST COAST FOREST	1261
WESTERN SIERRA MADRE PIEDMONT	55

Then, look at the spatial distribution and environmental characteristics of the grouped ecoregions

map1 <- ggplot() +
  geom_sf(data=cropped_states,fill='white') +
  geom_sf(data=modDat_1_sf#[modDat_1_sf$NA_L2NAME %in% c("MIXED WOOD PLAINS"),]
          ,
          aes(fill=as.factor(NA_L2NAME)),linewidth=0.5,alpha=0.5) +
  geom_point(data=modDat_1_s#[modDat_1_s$NA_L2NAME %in% c("MIXED WOOD PLAINS"),]
             ,
             alpha=0.5, 
             aes(x = Long, y = Lat, color=as.factor(NA_L2NAME)), alpha = .1) +
  theme(legend.position = 'none') +
  labs(title = "Level 2 Ecoregions as spatial blocks")

hull <- modDat_1_sf %>%
  ungroup() %>%
  group_by(NA_L2NAME) %>%
  slice(chull(tmean, prcp))

plot1<-ggplot(data=modDat_1_sf,aes(x=tmean,y=prcp)) +
  geom_polygon(data = hull, alpha = 0.25,aes(fill=NA_L2NAME) )+
  geom_point(aes(group=NA_L2NAME,color=NA_L2NAME),alpha=0.25) +
  theme_minimal() + xlab("Annual Average T_mean - long-term average") +
  ylab("Annual Average Precip - long-term average") 

plot2<-ggplot(data=modDat_1_sf %>%
                pivot_longer(cols=tmean:prcp),
              aes(x=value,group=name)) +
  # geom_polygon(data = hull, alpha = 0.25,aes(fill=fold) )+
  geom_density(aes(group=NA_L2NAME,fill=NA_L2NAME),alpha=0.25) +
  theme_minimal() +
  facet_wrap(~name,scales='free')
 
library(patchwork)
(combo <- (map1+plot1)/plot2) 

Fit a global model with all of the data

First, fit a LASSO regression model using the glmnet R package

• This regression is a Gamma glm with a log link
• Use cross validation across level 2 ecoregions to tune the lambda parameter in the LASSO model
• this model is fit to using the scaled weather/climate/soils variables
• this list of possible predictors includes:
  1. main effects
  2. interactions between all soils variables
  3. interactions between climate and weather variables
  4. transformed main effects (squared, log-transformed (add a uniform integer – 20– to all variables prior to log-transformation), square root-transformed (add a uniform integer – 20– to all variables prior to log-transformation))

Get rid of transformed predictions and interactions that are correlated

# get first pass of names correlated variables
X_df <- X %>% 
  as.data.frame() %>% 
  dplyr::select(-'(Intercept)')  
corrNames_i <- X_df %>% 
  cor()  %>% 
   caret::findCorrelation(cutoff = .7, verbose = FALSE, names = TRUE, exact = TRUE)
# remove those names that are untransformed main effects 
  # vector of columns to remove 
badNames <- corrNames_i[!(corrNames_i %in% prednames)]
while (sum(!(corrNames_i %in% prednames))>0) {
 corrNames_i <-  X_df %>% 
    dplyr::select(-badNames) %>% 
     cor()  %>% 
   caret::findCorrelation(cutoff = .7, verbose = FALSE, names = TRUE, exact = TRUE)
 # update the vector of names to remove 
 badNames <- unique(c(badNames, corrNames_i[!(corrNames_i %in% prednames)]))
}

## see if there are any correlated variables left (would be all main effects at this point)
# if there are, step through and remove the variable that each is most correlated with 
if (length(corrNames_i)>1) {
  for (i in 1:length(corrNames_i)) {
    X_i <- X_df %>% 
      dplyr::select(-badNames)
    if (corrNames_i[i] %in% names(X_i)) {
    corMat_i <- cor(x = X_i[corrNames_i[i]], y = X_i %>% dplyr::select(-corrNames_i[i])) 
    badNames_i <- colnames(corMat_i)[abs(corMat_i)>=.7]
    # if there are any predictors in the 'badNames_i', remove them from this list
    if (length(badNames_i) > 0 & sum(c(badNames_i %in% prednames))>0) {
        badNames_i <- badNames_i[!(badNames_i %in% prednames)]
    }
    badNames <- unique(c(badNames, badNames_i))
    }
  }
}
## update the X matrix to exclude these correlated variables
X <- X[,!(colnames(X) %in% badNames)]

if (autoKfold == FALSE) {
  # get the ecoregions for training lambda
  train_eco <- modDat_1_s$NA_L2NAME#[train]
  
  # Fit model -----------------------------------------------
  # specify leave-one-year-out cross-validation
  my_folds <- as.numeric(as.factor(train_eco))

  if (run == TRUE) {
    # set up parallel processing
    library(doMC)
    # this computer has 16 cores (parallel::detectCores())
    registerDoMC(cores = 8)
    
    fit <- cv.glmnet(
    x = X[,2:ncol(X)], 
    y = y, 
    family = stats::Gamma(link = "log"),
    keep = FALSE,
    alpha = 1,  # 0 == ridge regression, 1 == lasso, 0.5 ~~ elastic net
    #lambda = lambdas, 
    nlambda = 50,
    type.measure="mse",
    #penalty.factor = pen_facts,
    foldid = my_folds,
    #thresh = thresh,
    standardize = FALSE, ## scales variables prior to the model sequence... coefficients are always returned on the original scale
    parallel = TRUE#, 
    #relax = ifelse(response == "ShrubCover", yes = TRUE, no = FALSE)
    )
    base::saveRDS(fit, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/", response, "_globalLASSOmod_gammaLogLink_",ecoregion,".rds"))
  
  } else {
    fit <- readRDS(paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/", response, "_globalLASSOmod_gammaLogLink_",ecoregion, ".rds"))
  }
  
} else if (autoKfold == TRUE) {
   if (run == TRUE) {
    # set up parallel processing
    library(doMC)
    # this computer has 16 cores (parallel::detectCores())
    registerDoMC(cores = 8)
    
    fit <- cv.glmnet(
    x = X[,2:ncol(X)], 
    y = y, 
    family = stats::Gamma(link = "log"),
    keep = FALSE,
    alpha = .5,  # 0 == ridge regression, 1 == lasso, 0.5 ~~ elastic net
    lambda = lambdas,
    relax = ifelse(response == "ShrubCover", yes = TRUE, no = FALSE),
    #nlambda = 100,
    type.measure="mse",
    #penalty.factor = pen_facts,
    #foldid = my_folds,
    #thresh = thresh,
    standardize = FALSE, ## scales variables prior to the model sequence... coefficients are always returned on the original scale
    parallel = TRUE
    )
    base::saveRDS(fit, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/", response, "_globalLASSOmod_gammaLogLink_",ecoregion, "_kFoldDefault.rds"))
  
  } else {
    fit <- readRDS(paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/", response, "_globalLASSOmod_gammaLogLink_",ecoregion, "_kFoldDefault.rds"))
  }
}

  # assess model fit
  # assess.glmnet(fit$fit.preval, #newx = X[,2:293], 
  #               newy = y, family = stats::Gamma(link = "log"))
  # save the minimum lambda
  best_lambda <- fit$lambda.min
  # save the lambda for the most regularized model w/ an MSE that is still 1SE w/in the best lambda model
  lambda_1SE <- fit$lambda.1se
  # save the lambda for the most regularized model w/ an MSE that is still .5SE w/in the best lambda model
  lambda_halfSE <- best_lambda + ((lambda_1SE - best_lambda)/2)
 
  print(fit)     

## 
## Call:  cv.glmnet(x = X[, 2:ncol(X)], y = y, type.measure = "mse", foldid = my_folds,      keep = FALSE, parallel = TRUE, family = stats::Gamma(link = "log"),      alpha = 1, nlambda = 50, standardize = FALSE) 
## 
## Measure: Mean-Squared Error 
## 
##       Lambda Index Measure     SE Nonzero
## min 0.001105    27  0.5738 0.1094      25
## 1se 0.004970    19  0.6649 0.1224      18

  plot(fit)

Now, we need to do stability selection to ensure the coefficients that are being chosen with each lambda are stable

## stability selection for best lambda model 
# setup params
p <- ncol(X[,2:ncol(X)]) # of parameters
n <- length(y) # of observations
n_iter <- 100        # number of subsamples
sample_frac <- 0.75  # fraction of data to subsample
lambda_val <- best_lambda    # fixed lambda value (could be chosen via CV)

# Track selection
selection_counts <- matrix(0, nrow = p, ncol = 1)

for (i in 1:n_iter) {
  # Subsample rows
  sample_idx <- sample(1:n, size = floor(sample_frac * n), replace = FALSE)
  X_sub <- X[sample_idx, ]
  y_sub <- y[sample_idx]

  # Fit Lasso model
  fit_stab_i <- glmnet(x = X_sub[,2:ncol(X_sub)], y = y_sub, 
    family = stats::Gamma(link = "log"),
    alpha = 1, lambda = lambda_val, standardize = FALSE)

  # Get non-zero coefficients (excluding intercept)
  selected <- which(as.vector(coef(fit_stab_i))[-1] != 0)
  selection_counts[selected] <- selection_counts[selected] + 1
}

# Convert counts to selection probabilities (the probability that a variable is selected over 100 iterations)
selection_prob <- selection_counts / n_iter
selection_prob_df <- data.frame(
  VariableNumber = paste0("X", 1:p),
  VariableName = rownames(coef(fit_stab_i))[2:(p+1)],
  SelectionProb = as.vector(selection_prob)
)

# get those variables that are selected in ≥70% of subsets (Meinshausen and Bühlmann, 2010)
bestLambda_coef <- selection_prob_df[selection_prob_df$SelectionProb>=.7, c("VariableName", "SelectionProb")]

#//////
# stability selection for 1se lambda model
lambda_val <-  lambda_1SE    # fixed lambda value (could be chosen via CV)

# Track selection
selection_counts <- matrix(0, nrow = p, ncol = 1)

for (i in 1:n_iter) {
  # Subsample rows
  sample_idx <- sample(1:n, size = floor(sample_frac * n), replace = FALSE)
  X_sub <- X[sample_idx, ]
  y_sub <- y[sample_idx]

  # Fit Lasso model
  fit_stab_i <- glmnet(x = X_sub[,2:ncol(X_sub)], y = y_sub, 
    family = stats::Gamma(link = "log"),
    alpha = 1, lambda = lambda_val, standardize = FALSE)

  # Get non-zero coefficients (excluding intercept)
  selected <- which(as.vector(coef(fit_stab_i))[-1] != 0)
  selection_counts[selected] <- selection_counts[selected] + 1
}

# Convert counts to selection probabilities (the probability that a variable is selected over 100 iterations)
selection_prob <- selection_counts / n_iter
selection_prob_df <- data.frame(
  VariableNumber = paste0("X", 1:p),
  VariableName = rownames(coef(fit_stab_i))[2:(p+1)],
  SelectionProb = as.vector(selection_prob)
)

# get those variables that are selected in ≥70% of subsets (Meinshausen and Bühlmann, 2010)
seLambda_coef <- selection_prob_df[selection_prob_df$SelectionProb>=.7, c("VariableName", "SelectionProb")]

# #//////
# stability selection for half se lambda model
lambda_val <- lambda_halfSE    # fixed lambda value (could be chosen via CV)

# Track selection
selection_counts <- matrix(0, nrow = p, ncol = 1)

for (i in 1:n_iter) {
  # Subsample rows
  sample_idx <- sample(1:n, size = floor(sample_frac * n), replace = FALSE)
  X_sub <- X[sample_idx, ]
  y_sub <- y[sample_idx]

  # Fit Lasso model
  fit_stab_i <- glmnet(x = X_sub[,2:ncol(X_sub)], y = y_sub,
    family = stats::Gamma(link = "log"),
    alpha = 1, lambda = lambda_val, standardize = FALSE)

  # Get non-zero coefficients (excluding intercept)
  selected <- which(as.vector(coef(fit_stab_i))[-1] != 0)
  selection_counts[selected] <- selection_counts[selected] + 1
}

# Convert counts to selection probabilities (the probability that a variable is selected over 100 iterations)
selection_prob <- selection_counts / n_iter
selection_prob_df <- data.frame(
  VariableNumber = paste0("X", 1:p),
  VariableName = rownames(coef(fit_stab_i))[2:(p+1)],
  SelectionProb = as.vector(selection_prob)
)

# get those variables that are selected in ≥70% of subsets (Meinshausen and Bühlmann, 2010)
halfseLambda_coef <- selection_prob_df[selection_prob_df$SelectionProb>=.7, c("VariableName", "SelectionProb")]

If prompted to do so by the input arguments, remove any predictors that are for weather anomalies whose corresponding climate predictor is not included in the model

# if (trimAnom == TRUE) {
#   # get unique predictors
#   bestLamb_all <- bestLambda_coef$VariableName %>% #[str_detect(bestLambda_coef$VariableName, pattern = "_anom_s")] %>% 
#     str_remove(pattern = "I\\(") %>% 
#     str_remove(pattern = "\\^2\\)") %>% 
#     str_remove(pattern = "\\^2\\)") %>% 
#     str_split(pattern = ":", simplify = TRUE) 
#   if (ncol(bestLamb_all) == 1) {
#     bestLamb_all <- unique(bestLamb_all)
#   } else {
#     bestLamb_all <-  c(bestLamb_all[bestLamb_all[,1] !="",1], bestLamb_all[bestLamb_all[,2] !="",2]) %>% 
#       unique()
#   }
#   # get anomalies
#    bestLamb_anom <- bestLamb_all[bestLamb_all %in% prednames_weath]
#   # get climate
#    bestLamb_clim <- bestLamb_all[bestLamb_all %in% prednames_clim]
#   # which anomalies are present in the predictors, but their corresponding climate variable isn't
#    bestLamb_anomsWithMissingClim <- bestLamb_anom[!(str_remove(bestLamb_anom, "_anom") %in% bestLamb_clim)]
#    # remove anomalies (and all interaction terms w/ those anomalies) from the predictor list
# 
#    if (length(bestLamb_anomsWithMissingClim) != 0) {
#    
#         bestLambda_coef_NEW <- bestLambda_coef
#       for (i in 1:length(bestLamb_anomsWithMissingClim)) {
#      bestLambda_coef_NEW <- bestLambda_coef_NEW[!str_detect(bestLambda_coef_NEW$VariableName, pattern = bestLamb_anomsWithMissingClim[i]),]
#       }
#      
#    bestLambda_coef <- bestLambda_coef_NEW
#    }
#    
#    
#   #//// 1 se lambda model
#       if (nrow(seLambda_coef) !=0) {
#         # get unique predictors
#   seLamb_all <- seLambda_coef$VariableName %>% #[str_detect(seLambda_coef$VariableName, pattern = "_anom_s")] %>% 
#     str_remove(pattern = "I\\(") %>% 
#     str_remove(pattern = "_s\\^2\\)") %>% 
#     str_remove(pattern = "\\^2\\)") %>% 
#     str_split(pattern = ":", simplify = TRUE) 
#   if (ncol(seLamb_all) == 1) {
#     seLamb_all <- unique(seLamb_all)
#   } else {
#     seLamb_all <-  c(seLamb_all[seLamb_all[,1] !="",1], seLamb_all[seLamb_all[,2] !="",2]) %>% 
#       unique()
#   }
#   # get anomalies
#    seLamb_anom <- seLamb_all[seLamb_all %in% prednames_weath]
#   # get climate
#    seLamb_clim <- seLamb_all[seLamb_all %in% prednames_clim]
#   # which anomalies are present in the predictors, but their corresponding climate variable isn't
#    seLamb_anomsWithMissingClim <- seLamb_anom[!(str_remove(seLamb_anom, "_anom") %in%  seLamb_clim)]
#    # remove anomalies (and all interaction terms w/ those anomalies) from the predictor list
#       if (length(seLamb_anomsWithMissingClim) != 0) {
#     seLambda_coef_NEW <- seLambda_coef
#    for (i in 1:length(seLamb_anomsWithMissingClim)) {
#      seLambda_coef_NEW <- seLambda_coef_NEW[!str_detect(seLambda_coef_NEW$VariableName, pattern = seLamb_anomsWithMissingClim[i]),]
#    }
#    seLambda_coef <- seLambda_coef_NEW
#       }
#    }
#   
#   #//// 1 se lambda model
#       if (nrow(halfseLambda_coef) !=0) {
#         # get unique predictors
#   halfseLamball<- halfseLambda_coef$VariableName %>% #[str_detect(halfseLambda_coef$VariableName, pattern = "_anom_s")] %>% 
#     str_remove(pattern = "I\\(") %>% 
#     str_remove(pattern = "_s\\^2\\)") %>% 
#     str_split(pattern = ":", simplify = TRUE) 
#   
#     if (ncol(halfseLamball) == 1) {
#     halfseLamball <- unique(halfseLamball)
#   } else {
#     halfseLamball <-  c(halfseLamball[halfseLamball[,1] !="",1], halfseLamball[halfseLamball[,2] !="",2]) %>% 
#       unique()
#   }
#   
#   # get anomalies
#    halfseLambanom <- halfseLamball[halfseLamball %in% prednames_weath]
#   # get climate
#    halfseLambclim <- halfseLamball[halfseLamball %in% prednames_clim]
#   # which anomalies are present in the predictors, but their corresponding climate variable isn't
#    halfseLambanomsWithMissingClim <- halfseLambanom[!(str_remove(halfseLambanom, "_anom_s") %in%  halfseLambclim)]
#    # remove anomalies (and all interaction terms w/ those anomalies) from the predictor list
# 
#    
#    ##
#         if (length(halfseLambanomsWithMissingClim) != 0) {
#    halfseLambda_coef_NEW <- halfseLambda_coef
#    for (i in 1:length(halfseLambanomsWithMissingClim)) {
#      halfseLambda_coef_NEW <- halfseLambda_coef_NEW[!str_detect(halfseLambda_coef_NEW$VariableName, pattern = halfseLambanomsWithMissingClim[i]),]
#    }
#    halfseLambda_coef <- halfseLambda_coef_NEW
#       }
#    }
#     ##
#   
#       }

Then, fit regular glm models (Gamma glm with a log link), first using the coefficients from the ‘best’ lambda identified in the LASSO model, and then using the coefficients from the ‘1SE’ lambda and then the ‘1/2SE’ lambda values identified from the LASSO (this is the value of lambda such that the cross validation error is within 1 standard error of the minimum).

## fit w/ the identified coefficients from the 'best' lambda, but using the glm function
  mat_glmnet_best <- bestLambda_coef$VariableName 

  if (length(mat_glmnet_best) == 0) {
    f_glm_bestLambda <- as.formula(paste0("response_transformed ~  1"))
  } else {
  f_glm_bestLambda <- as.formula(paste0( "response_transformed ~ ", paste0(mat_glmnet_best, collapse = " + ")))
  }
  
  fit_glm_bestLambda <- glm(data = modDat_1_s
                              , formula =  f_glm_bestLambda, family =  stats::Gamma(link = "log"))
  
   ## fit w/ the identified coefficients from the '1se' lambda, but using the glm function
  mat_glmnet_1se <- seLambda_coef$VariableName

  if (length(mat_glmnet_1se) == 0) {
    f_glm_1se <- as.formula(paste0("response_transformed ~  1"))
  } else {
  f_glm_1se <- as.formula(paste0("response_transformed ~ ", paste0(mat_glmnet_1se, collapse = " + ")))
  }


  fit_glm_se <- glm(data = modDat_1_s, formula = f_glm_1se,
                    family =  stats::Gamma(link = "log"))

     ## fit w/ the identified coefficients from the '.5se' lambda, but using the glm function
  mat_glmnet_halfse <- halfseLambda_coef$VariableName

  if (length(mat_glmnet_halfse) == 0) {
    f_glm_halfse <- as.formula(paste0(response, "response_transformed ~  1"))
  } else {
  f_glm_halfse <- as.formula(paste0("response_transformed ~" , paste0(mat_glmnet_halfse, collapse = " + ")))
  }

  fit_glm_halfse <- glm(data = modDat_1_s, formula =  f_glm_halfse,
                    family =  stats::Gamma(link = "log"))

  ## save models 
  if (trimAnom == TRUE) {
    saveRDS(fit_glm_bestLambda, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/",response,"_",ecoregion, "_bestLambdaGLM.rds"))
  saveRDS(fit_glm_halfse, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/",response,"_",ecoregion, "_halfSELambdaGLM.rds"))
  saveRDS(fit_glm_se, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/",response,"_",ecoregion, "_oneSELambdaGLM.rds"))
  } else {
    saveRDS(fit_glm_bestLambda, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/",response,"_",ecoregion, "_noTLP_",removeTLP,"_bestLambdaGLM.rds"))
  saveRDS(fit_glm_halfse, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/",response,"_",ecoregion, "_noTLP_",removeTLP,"_halfSELambdaGLM.rds"))
  saveRDS(fit_glm_se, paste0("/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Analysis/BiomassQuantity/Analysis/models/",response,"_",ecoregion, "_noTLP_",removeTLP,"_oneSELambdaGLM.rds"))
    
  }

Then, we predict (on the training set) using both of these models (best lambda and 1se lambda)

  ## predict on the test data
  # lasso model predictions with the optimal lambda
  optimal_pred <- predict(fit_glm_bestLambda, newx=X[,2:ncol(X)], type = "response") -2
  optimal_pred_1se <-  predict(fit_glm_se, newx=X[,2:ncol(X)], type = "response") -2
  optimal_pred_halfse <- predict(fit_glm_halfse, newx = X[,2:ncol(X)], type = "response") -2
  
    null_fit <- glm(#data = data.frame("y" = y, X[,paste0(prednames, "_s")]), 
      formula = y ~ 1, family = stats::Gamma(link = "log"))
  null_pred <- predict(null_fit, newdata = as.data.frame(X), type = "response"
                       ) -2

  # ## snap values above 100 and below 0 to be 100 or z
  # #optimal_pred[optimal_pred>100] <- 100
  # optimal_pred[optimal_pred<0] <- 0
  # #optimal_pred_1se[optimal_pred_1se>100] <- 100
  # optimal_pred_1se[optimal_pred_1se<0] <- 0
  # #optimal_pred_halfse[optimal_pred_halfse>100] <- 100
  # optimal_pred_halfse[optimal_pred_halfse<0] <- 0
  # 
  # save data
  fullModOut <- list(
    "modelObject" = fit,
    "nullModelObject" = null_fit,
    "modelPredictions" = data.frame(#ecoRegion_holdout = rep(test_eco,length(y)),
      obs=y-2,
                    pred_opt=optimal_pred, 
                    pred_opt_se = optimal_pred_1se,
                    pred_opt_halfse = optimal_pred_halfse,
                    pred_null=null_pred#,
                    #pred_nopenalty=nopen_pred
                    ))
  
# # calculate correlations between null and optimal model 
# my_cors <- c(cor(optimal_pred, c(y)),
#              cor(optimal_pred_1se, c(y)), 
#             cor(null_pred, c(y))
#             )
# 
# # calculate mse between null and optimal model 
# my_mse <- c(mean((fullModOut$modelPredictions$pred_opt -  c(y))^2) ,
#             mean((fullModOut$modelPredictions$pred_opt_se -  c(y))^2) ,
#             mean((fullModOut$modelPredictions$pred_null - c(y))^2)#,
#             #mean((obs_pred$pred_nopenalty - obs_pred$obs)^2)
#             )

ggplot() + 
  geom_point(aes(X[,2], fullModOut$modelPredictions$obs), col = "black", alpha = .1) + 
  geom_point(aes(X[,2], fullModOut$modelPredictions$pred_opt), col = "red", alpha = .1) + ## predictions w/ the CV model
  geom_point(aes(X[,2], fullModOut$modelPredictions$pred_opt_halfse), col = "orange", alpha = .1) + ## predictions w/ the CV model (.5se lambda)
  geom_point(aes(X[,2], fullModOut$modelPredictions$pred_opt_se), col = "green", alpha = .1) + ## predictions w/ the CV model (1se lambda)
  geom_point(aes(X[,2], fullModOut$modelPredictions$pred_null), col = "blue", alpha = .1) + 
  labs(title = "A rough comparison of observed and model-predicted values", 
       subtitle = "black = observed values \n red = predictions from 'best lambda' model \n orange = predictions for '1/2se' lambda model \n green = predictions from '1se' lambda model \n blue = predictions from null model") +
  xlab(colnames(X)[2])

  #ylim(c(0,200))

The internal cross-validation process to fit the global LASSO model identified an optimal lambda value (regularization parameter) of r{print(best_lambda)}. The lambda value such that the cross validation error is within 1 standard error of the minimum (“1se lambda”) was `r{print(fit$lambda.1se)}`` . The following coefficients were kept in each model:

# the coefficient matrix from the 'best model' -- find and print those coefficients that aren't 0 in a table
coef_glm_bestLambda <- coef(fit_glm_bestLambda) %>% 
  data.frame() 
coef_glm_bestLambda$coefficientName <- rownames(coef_glm_bestLambda)
names(coef_glm_bestLambda)[1] <- "coefficientValue_bestLambda"
# coefficient matrix from the '1se' model 
coef_glm_1se <- coef(fit_glm_se) %>% 
  data.frame() 
coef_glm_1se$coefficientName <- rownames(coef_glm_1se)
names(coef_glm_1se)[1] <- "coefficientValue_1seLambda"
# coefficient matrix from the 'half se' model 
coef_glm_halfse <- coef(fit_glm_halfse) %>% 
  data.frame() 
coef_glm_halfse$coefficientName <- rownames(coef_glm_halfse)
names(coef_glm_halfse)[1] <- "coefficientValue_halfseLambda"
# add together
coefs <- full_join(coef_glm_bestLambda, coef_glm_halfse) %>% 
  full_join(coef_glm_1se) %>% 
  dplyr::select(coefficientName, coefficientValue_bestLambda,
                coefficientValue_halfseLambda, coefficientValue_1seLambda)

globModTerms <- coefs[!is.na(coefs$coefficientValue_bestLambda), "coefficientName"]

## also, get the number of unique variables in each model 
var_prop_pred <- paste0("biomass_MgPerHect", "_pred")
response_vars <- c("biomass_MgPerHect", var_prop_pred)
# for best lambda model
prednames_fig <- paste(str_split(globModTerms, ":", simplify = TRUE)) 
prednames_fig <- str_replace(prednames_fig, "I\\(", "")
prednames_fig <- str_replace(prednames_fig, "\\^2\\)", "")
prednames_fig <- unique(prednames_fig[prednames_fig>0])
prednames_fig <- prednames_fig
prednames_fig_num <- length(prednames_fig)
# for 1SE lambda model
globModTerms_1se <- coefs[!is.na(coefs$coefficientValue_1seLambda), "coefficientName"]
if (length(globModTerms_1se) == 1) {
prednames_fig_1se <- paste(str_split(globModTerms_1se, ":", simplify = TRUE)) 
prednames_fig_1se <- str_replace(prednames_fig_1se, "I\\(", "")
prednames_fig_1se <- str_replace(prednames_fig_1se, "\\^2\\)", "")
prednames_fig_1se <- unique(prednames_fig_1se[prednames_fig_1se>0])
prednames_fig_1se_num <- c(0)
} else {
prednames_fig_1se <- paste(str_split(globModTerms_1se, ":", simplify = TRUE)) 
prednames_fig_1se <- str_replace(prednames_fig_1se, "I\\(", "")
prednames_fig_1se <- str_replace(prednames_fig_1se, "\\^2\\)", "")
prednames_fig_1se <- unique(prednames_fig_1se[prednames_fig_1se>0])
prednames_fig_1se_num <- length(prednames_fig_1se)
}
# for 1/2SE lambda model
globModTerms_halfse <- coefs[!is.na(coefs$coefficientValue_halfseLambda), "coefficientName"]
if (length(globModTerms_halfse) == 1) {
prednames_fig_halfse <- paste(str_split(globModTerms_halfse, ":", simplify = TRUE)) 
prednames_fig_halfse <- str_replace(prednames_fig_halfse, "I\\(", "")
prednames_fig_halfse <- str_replace(prednames_fig_halfse, "\\^2\\)", "")
prednames_fig_halfse <- unique(prednames_fig_halfse[prednames_fig_halfse>0])
prednames_fig_halfse_num <- c(0)
} else {
prednames_fig_halfse <- paste(str_split(globModTerms_halfse, ":", simplify = TRUE)) 
prednames_fig_halfse <- str_replace(prednames_fig_halfse, "I\\(", "")
prednames_fig_halfse <- str_replace(prednames_fig_halfse, "\\^2\\)", "")
prednames_fig_halfse <- unique(prednames_fig_halfse[prednames_fig_halfse>0])
prednames_fig_halfse_num <- length(prednames_fig_halfse)
}
# make a table
kable(coefs, col.names = c("Coefficient Name", "Value from best lambda model", 
                           "Value form 1/2 se lambda", "Value from 1se lambda model")
      ) %>%
kable_styling(bootstrap_options = c("striped", "hover", "condensed")) 

Coefficient Name	Value from best lambda model	Value form 1/2 se lambda	Value from 1se lambda model
(Intercept)	1.2143987	1.2195449	1.2199574
isothermality	0.0143111	NA	NA
prcp_seasonality	0.0205236	0.0238948	NA
annWetDegDays	0.1462042	0.1447334	0.1338101
t_warm	0.0645604	0.0605909	0.0671457
t_cold	-0.0026725	0.0148926	0.0212248
prcp_wet	0.0609698	0.0565644	0.0620226
sand	-0.0447523	-0.0296785	NA
coarse	-0.0263447	-0.0204261	-0.0097440
AWHC	-0.0627533	-0.0650574	-0.0099693
clay	-0.0176512	NA	NA
I(prcpTempCorr^2)	0.0226938	0.0249986	0.0222020
I(isothermality^2)	-0.0365198	-0.0363227	-0.0329701
I(prcp_wet^2)	-0.0624455	-0.0601663	-0.0645850
I(sand^2)	-0.0045461	-0.0052849	NA
isothermality:t_warm	-0.0402761	NA	NA
prcp_wet:prcpTempCorr	-0.0023931	-0.0064876	-0.0107446
t_cold:prcp_wet	0.0500153	0.0451696	0.0494647
t_cold:prcpTempCorr	-0.0375529	-0.0319602	-0.0374593
t_warm:tmean	0.0120733	NA	NA
AWHC:soilDepth	-0.0453524	NA	NA
clay:carbon	0.0047739	NA	NA
coarse:clay	0.0198089	0.0250370	0.0156480
coarse:soilDepth	-0.0053201	NA	NA
sand:soilDepth	0.0291997	NA	NA
soilDepth	NA	0.0086169	-0.0380114
t_warm:isothermality	NA	-0.0360547	-0.0327831
soilDepth:AWHC	NA	-0.0429539	-0.0450356
soilDepth:sand	NA	0.0265255	0.0330239
I(clay^2)	NA	NA	0.0011337

# calculate RMSE of all models 
RMSE_best <- yardstick::rmse(fullModOut$modelPredictions[,c("obs", "pred_opt")], truth = "obs", estimate = "pred_opt")$.estimate
RMSE_halfse <- yardstick::rmse(fullModOut$modelPredictions[,c("obs", "pred_opt_halfse")], truth = "obs", estimate = "pred_opt_halfse")$.estimate
RMSE_1se <- yardstick::rmse(fullModOut$modelPredictions[,c("obs", "pred_opt_se")], truth = "obs", estimate = "pred_opt_se")$.estimate
# calculate bias of all models
bias_best <-  mean((fullModOut$modelPredictions$obs) - fullModOut$modelPredictions$pred_opt)
bias_halfse <-  mean((fullModOut$modelPredictions$obs) - fullModOut$modelPredictions$pred_opt_halfse)
bias_1se <- mean((fullModOut$modelPredictions$obs) - fullModOut$modelPredictions$pred_opt_se)

uniqueCoeffs <- data.frame("Best lambda model" = c(RMSE_best, bias_best,
  as.integer(length(globModTerms)-1), as.integer(prednames_fig_num), 
                                                   as.integer(sum(prednames_fig %in% c(prednames_clim))),
                                                   #as.integer(sum(prednames_fig %in% c(prednames_weath))),
                                                   as.integer(sum(prednames_fig %in% c(prednames_soils)))
                                                   ), 
                           "1/2 se lambda model" = c(RMSE_halfse, bias_halfse,
                             length(globModTerms_halfse)-1, prednames_fig_halfse_num,
                                                   sum(prednames_fig_halfse %in% c(prednames_clim)),
                                                   #sum(prednames_fig_halfse %in% c(prednames_weath)),
                                                   sum(prednames_fig_halfse %in% c(prednames_soils))), 
                           "1se lambda model" = c(RMSE_1se, bias_1se,
                             length(globModTerms_1se)-1, prednames_fig_1se_num,
                                                   sum(prednames_fig_1se %in% c(prednames_clim)),
                                                   #sum(prednames_fig_1se %in% c(prednames_weath)),
                                                   sum(prednames_fig_1se %in% c(prednames_soils))))
row.names(uniqueCoeffs) <- c("RMSE", "bias: mean(obs-pred.)", "Total number of coefficients", "Number of unique coefficients",
                             "Number of unique climate coefficients", 
                             "Number of unique soils coefficients"
                             )

kable(uniqueCoeffs, 
      col.names = c("Best lambda model", "1/2 se lambda model", "1se lambda model"), row.names = TRUE) %>%
kable_styling(bootstrap_options = c("striped", "hover", "condensed")) 

	Best lambda model	1/2 se lambda model	1se lambda model
RMSE	0.6232563	0.6244400	0.6262582
bias: mean(obs-pred.)	0.0007872	0.0008798	0.0007364
Total number of coefficients	24.0000000	20.0000000	18.0000000
Number of unique coefficients	14.0000000	12.0000000	11.0000000
Number of unique climate coefficients	8.0000000	7.0000000	6.0000000
Number of unique soils coefficients	6.0000000	5.0000000	5.0000000

Visualizations of Model Predictions and Residuals – using best lambda model

observed vs. predicted values

If the 1se lambda has no terms (is an intercept only model), use the 1/2 se lambda in subsequent figures

if (length(prednames_fig_1se) == 0) {
  mod_secondBest <- fit_glm_halfse
  name_secondBestMod <- "1/2 SE Model"
  prednames_secondBestMod <- prednames_fig_halfse
} else {
  mod_secondBest <- fit_glm_se
  name_secondBestMod <- "1 SE Model"
  prednames_secondBestMod <- prednames_fig_1se
}

Predicting on the data

  # create prediction for each each model
# (i.e. for each fire proporation variable)
predict_by_response <- function(mod, df) {
  df_out <- df
  response_name <- paste0("biomass_MgPerHect", "_pred")
  preds <- predict(mod, newx= df_out, #s="lambda.min", 
                                     type = "response")
  preds <- preds -2
  preds[preds<0] <- 0
  #preds[preds>100] <- 100
  df_out <- df_out %>% cbind(preds)
   colnames(df_out)[ncol(df_out)] <- response_name
  return(df_out)
}

pred_glm1 <- predict_by_response(fit_glm_bestLambda, X[,2:ncol(X)])

## back-transform the 
# add back in true y values
pred_glm1 <- pred_glm1 %>% 
  cbind(modDat_1_s[,c("biomass_MgPerHect", "response_transformed")])

# add back in lat/long data 
pred_glm1 <- pred_glm1 %>% 
  cbind(modDat_1_s[,c("Long", "Lat")])

# calculate residuals
pred_glm1$resid <- pred_glm1[,"biomass_MgPerHect"] - pred_glm1[,paste0("biomass_MgPerHect", "_pred")]
pred_glm1$extremeResid <- NA
pred_glm1[pred_glm1$resid > 70 | pred_glm1$resid < -70,"extremeResid"] <- 1

# calculate residuals as percentages of the total 
pred_glm1$resid_perc <- (pred_glm1$resid / pred_glm1[,"biomass_MgPerHect"]) * 100

# plot(x = pred_glm1[,response],
#      y = pred_glm1[,paste0("biomass_MgPerHect", "_pred")],
#      xlab = "observed values", ylab = "predicted values")
# points(x = pred_glm1[!is.na(pred_glm1$extremeResid), response],
#        y = pred_glm1[!is.na(pred_glm1$extremeResid), paste0("biomass_MgPerHect", "_pred")],
#        col = "red")

pred_glm1_1se <- predict_by_response(mod_secondBest, X[,2:ncol(X)])

# add back in true y values
pred_glm1_1se <- pred_glm1_1se %>% 
  cbind(modDat_1_s[,c("biomass_MgPerHect", "response_transformed")])

# add back in lat/long data 
pred_glm1_1se <- pred_glm1_1se %>% 
  cbind(modDat_1_s[,c("Long", "Lat")])

# calculate residuals
pred_glm1_1se$resid <- pred_glm1_1se[,"biomass_MgPerHect"] - pred_glm1_1se[,paste0("biomass_MgPerHect", "_pred")]
pred_glm1_1se$extremeResid <- NA
pred_glm1_1se[pred_glm1_1se$resid > 70 | pred_glm1_1se$resid < -70,"extremeResid"] <- 1

# calculate residuals as percentages of the total 
pred_glm1_1se$resid_perc <- (pred_glm1_1se$resid / pred_glm1_1se[,"biomass_MgPerHect"]) * 100

Maps of Observations, Predictions, and Residuals=

Observations across the temporal range of the dataset

pred_glm1 <- pred_glm1 %>% 
  mutate(resid = .[["biomass_MgPerHect"]] - .[[paste0("biomass_MgPerHect","_pred")]]) 

# rasterize
# get reference raster
test_rast <-  rast("../../../Data_raw/dayMet/rawMonthlyData/orders/70e0da02b9d2d6e8faa8c97d211f3546/Daymet_Monthly_V4R1/data/daymet_v4_prcp_monttl_na_1980.tif") %>% 
  terra::aggregate(fact = 8, fun = "mean")

## |---------|---------|---------|---------|=========================================                                          

## add ecoregion boundaries (for our ecoregion level model)
regions <- sf::st_read(dsn = "../../../Data_raw/Level2Ecoregions/", layer = "NA_CEC_Eco_Level2") 

## Reading layer `NA_CEC_Eco_Level2' from data source 
##   `/Users/astears/Documents/Dropbox_static/Work/NAU_USGS_postdoc/PED_vegClimModels/Data_raw/Level2Ecoregions' 
##   using driver `ESRI Shapefile'
## Simple feature collection with 2261 features and 8 fields
## Geometry type: POLYGON
## Dimension:     XY
## Bounding box:  xmin: -4334052 ymin: -3313739 xmax: 3324076 ymax: 4267265
## Projected CRS: Sphere_ARC_INFO_Lambert_Azimuthal_Equal_Area

regions <- regions %>% 
  st_transform(crs = st_crs(test_rast)) %>% 
  st_make_valid() #%>% 
  #st_crop(st_bbox(test_rast))
# 
# goodRegions_temp <- st_overlaps(y = cropped_states, x = regions, sparse = FALSE) %>% 
#   rowSums() 
# goodRegions <- regions[goodRegions_temp != 0,]

ecoregionLU <- data.frame("NA_L1NAME" = sort(unique(regions$NA_L1NAME)), 
                        "newRegion" = c(NA, "Forest", "dryShrubGrass", 
                                        "dryShrubGrass", "Forest", "dryShrubGrass",
                                       "dryShrubGrass", "Forest", "Forest", 
                                       "dryShrubGrass", "Forest", "Forest", 
                                       "Forest", "Forest", "dryShrubGrass", 
                                       NA
                                        ))
goodRegions <- regions %>% 
  left_join(ecoregionLU)
mapRegions <- goodRegions %>% 
  filter(!is.na(newRegion)) %>% 
  group_by(newRegion) %>% 
  summarise(geometry = sf::st_union(geometry)) %>% 
  ungroup() %>% 
  st_simplify(dTolerance = 1000)
#mapview(mapRegions)
# rasterize data
plotObs <- pred_glm1 %>% 
         drop_na(paste("biomass_MgPerHect")) %>% 
  #slice_sample(n = 5e4) %>%
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) %>% 
  terra::rasterize(y = test_rast, 
                   field = "biomass_MgPerHect", 
                   fun = mean) #%>% 
  #terra::aggregate(fact = 2, fun = mean, na.rm = TRUE) %>% 
  #terra::crop(ext(-1950000, 1000000, -1800000, 1000000))

# get the extent of this particular raster, and crop it accordingly
tempExt <- crds(plotObs, na.rm = TRUE)

plotObs_2 <- plotObs %>% 
  crop(ext(min(tempExt[,1]), max(tempExt[,1]),
           min(tempExt[,2]), max(tempExt[,2])) 
       )
# make figures
map_obs <- ggplot() +
geom_spatraster(data = plotObs_2) + 
  geom_sf(data=cropped_states %>% st_transform(crs = st_crs(test_rast)) %>% st_crop(st_bbox(plotObs_2)),fill=NA ) +
  geom_sf(data = mapRegions, fill = NA, col = "orchid", lwd = .5) +
labs(title = paste0("Observations of ", "biomass_MgPerHect", " in the ",ecoregion, " ecoregion")) +
  scale_fill_gradient2(low = "brown",
                       mid = "wheat" ,
                       high = "darkgreen" , 
                       midpoint = 0,   na.value = "grey20") + 
  xlim(st_bbox(plotObs_2)[c(1,3)]) + 
  ylim(st_bbox(plotObs_2)[c(2,4)])

hist_obs <- ggplot(pred_glm1) + 
  geom_histogram(aes(.data[["biomass_MgPerHect"]]), fill = "lightgrey", col = "darkgrey")

library(ggpubr)
ggarrange(map_obs, hist_obs, heights = c(3,1), ncol = 1)

Predictions across the temporal range of the dataset

# rasterize data
plotPred <- pred_glm1 %>% 
         drop_na(paste0("biomass_MgPerHect","_pred")) %>% 
  #slice_sample(n = 5e4) %>%
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) %>% 
  terra::rasterize(y = test_rast, 
                   field = paste0("biomass_MgPerHect","_pred"), 
                   fun = mean) #%>% 
  #terra::aggregate(fact = 2, fun = mean, na.rm = TRUE) %>% 
  #terra::crop(ext(-1950000, 1000000, -1800000, 1000000))

# get the point location of those predictions that are > 600
highPred_points <- pred_glm1 %>% 
  filter(.[[paste0("biomass_MgPerHect","_pred")]] > 700 | 
           .[[paste0("biomass_MgPerHect", "_pred")]] < 0) %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 

# get the extent of this particular raster, and crop it accordingly
tempExt <- crds(plotPred, na.rm = TRUE)

plotPred_2 <- plotPred %>% 
  crop(ext(min(tempExt[,1]), max(tempExt[,1]),
           min(tempExt[,2]), max(tempExt[,2])) 
       )
 # make figures
map_preds1 <- ggplot() +
geom_spatraster(data = plotPred_2) + 
  geom_sf(data=cropped_states %>% st_transform(crs = st_crs(test_rast)) %>% st_crop(st_bbox(plotObs_2)),fill=NA )  + 
  geom_sf(data = mapRegions, fill = NA, col = "orchid", lwd = .5) +
  geom_sf(data = highPred_points, col = "red") +
labs(title = paste0("Predictions from the 'best lambda' fitted model of ", "biomass_MgPerHect", " in the ",ecoregion, " ecoregion"),
     subtitle =  "bestLambda model\n 
     red points show predictions greater than 700 Mg/hect")  +
  scale_fill_gradient2(low = "wheat",
                       high = "darkgreen" , 
                       midpoint = 0,   
                       na.value = "grey20",
                       limits = c(0,max(terra::values(plotPred), na.rm = TRUE))) + 
  xlim(st_bbox(plotObs_2)[c(1,3)]) + 
  ylim(st_bbox(plotObs_2)[c(2,4)])

hist_preds1 <- ggplot(pred_glm1) + 
  geom_histogram(aes(.data[[paste0("biomass_MgPerHect", "_pred")]]), fill = "lightgrey", col = "darkgrey")#+ 
  #xlim(c(0,700))

ggarrange(map_preds1, hist_preds1, heights = c(3,1), ncol = 1)

# rasterize data
plotPred <- pred_glm1_1se %>% 
         drop_na(paste0("biomass_MgPerHect","_pred")) %>% 
  #slice_sample(n = 5e4) %>%
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) %>% 
  terra::rasterize(y = test_rast, 
                   field = paste0("biomass_MgPerHect","_pred"), 
                   fun = mean) #%>% 
  #terra::aggregate(fact = 2, fun = mean, na.rm = TRUE) %>% 
  #terra::crop(ext(-1950000, 1000000, -1800000, 1000000))

# get the point location of those predictions that are > 100
highPred_points <- pred_glm1_1se %>% 
  filter(.[[paste0("biomass_MgPerHect","_pred")]] > 700 | 
           .[[paste0("biomass_MgPerHect", "_pred")]] < 0) %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 

# get the extent of this particular raster, and crop it accordingly
tempExt <- crds(plotPred, na.rm = TRUE)

plotPred_2 <- plotPred %>% 
  crop(ext(min(tempExt[,1]), max(tempExt[,1]),
           min(tempExt[,2]), max(tempExt[,2])) 
       )
# make figures
map_preds2 <- ggplot() +
geom_spatraster(data = plotPred_2) + 
  geom_sf(data=cropped_states %>% st_transform(crs = st_crs(test_rast)) %>% st_crop(st_bbox(plotObs_2)),fill=NA )  + 
  geom_sf(data = mapRegions, fill = NA, col = "orchid", lwd = .5) +
  geom_sf(data = highPred_points, col = "red") +
labs(title = paste0("Predictions from the '1SE lambda' fitted model of ", "biomass_MgPerHect", " in the ",ecoregion, " ecoregion"),
     subtitle =  name_secondBestMod)  +
  scale_fill_gradient2(low = "wheat",
                       high = "darkgreen" , 
                       #midpoint = 100,   
                       na.value = "grey20",
                       limits = c(0,max(terra::values(plotPred), na.rm = TRUE))) + 
  xlim(st_bbox(plotObs_2)[c(1,3)]) + 
  ylim(st_bbox(plotObs_2)[c(2,4)])

hist_preds2 <- ggplot(pred_glm1_1se) + 
  geom_histogram(aes(.data[[paste0("biomass_MgPerHect", "_pred")]]), fill = "lightgrey", col = "darkgrey")#+ 
  #xlim(c(0,200))

ggarrange(map_preds2, hist_preds2, heights = c(3,1), ncol = 1)

Residuals as a percentage of prediction ((resid - prediction)*100) for best SE lambda model

# rasterize data
plotResid_rast <- pred_glm1 %>% 
         drop_na(resid_perc) %>% 
  #slice_sample(n = 5e4) %>%
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) %>% 
  terra::rasterize(y = test_rast, 
                   field = "resid_perc", 
                   fun = mean) #%>% 
  #terra::aggregate(fact = 2, fun = mean, na.rm = TRUE) %>% 
  #terra::crop(ext(-1950000, 1000000, -1800000, 1000000))

# get the extent of this particular raster, and crop it accordingly
tempExt <- crds(plotResid_rast, na.rm = TRUE)

plotResid_rast_2 <- plotResid_rast %>% 
  crop(ext(min(tempExt[,1]), max(tempExt[,1]),
           min(tempExt[,2]), max(tempExt[,2])) 
       )

# identify locations where residuals are >100 or < -100
badResids_high <- pred_glm1 %>% 
  filter(resid > 100)  %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 
badResids_low <- pred_glm1 %>% 
  filter(resid < -100)  %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 
# make figures
map <- ggplot() +
geom_spatraster(data =plotResid_rast_2) + 
  geom_sf(data=cropped_states %>% st_transform(crs = st_crs(test_rast)) %>% st_crop(st_bbox(plotObs_2)),fill=NA )  + 
  geom_sf(data = mapRegions, fill = NA, col = "orchid", lwd = .5) +
  geom_sf(data = badResids_high, col = "blue") +
  geom_sf(data = badResids_low, col = "red") +
labs(title = paste0("Resids. as a percentage (obs. - pred.)/obs. * 100 from the ", ecoregion," ecoregion-wide model of ",  "biomass_MgPerHect"),
     subtitle = "bestLambda model \n red points indicate locations that have residuals below -200% \n blue points indicate locations that have residuals above 200%") +
  scale_fill_gradient2(low = "red",
                       mid = "white" ,
                       high = "blue" , 
                       midpoint = 0,   na.value = "grey20",
                       limits = c(-100 ,100)
                       ) + 
  xlim(st_bbox(plotObs_2)[c(1,3)]) + 
  ylim(st_bbox(plotObs_2)[c(2,4)])
hist <- ggplot(pred_glm1) + 
  geom_histogram(aes(resid), fill = "lightgrey", col = "darkgrey") + 
  geom_text(aes(x = min(resid)*.9, y = 1500, label = paste0("min = ", round(min(resid),2)))) +
  geom_text(aes(x = max(resid)*.9, y = 1500, label = paste0("max = ", round(max(resid),2)))) + 
  geom_vline(aes(xintercept = mean(resid)))

ggarrange(map, hist, heights = c(3,1), ncol = 1)

Residuals as a percentage of prediction ((resid - prediction)*100) for second best SE lambda model

# rasterize data
plotResid_rast <- pred_glm1_1se %>% 
         drop_na(resid_perc) %>% 
  #slice_sample(n = 5e4) %>%
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) %>% 
  terra::rasterize(y = test_rast, 
                   field = "resid_perc", 
                   fun = mean) #%>% 
  #terra::aggregate(fact = 2, fun = mean, na.rm = TRUE) %>% 
  #terra::crop(ext(-1950000, 1000000, -1800000, 1000000))

# get the extent of this particular raster, and crop it accordingly
tempExt <- crds(plotResid_rast, na.rm = TRUE)

plotResid_rast_2 <- plotResid_rast %>% 
  crop(ext(min(tempExt[,1]), max(tempExt[,1]),
           min(tempExt[,2]), max(tempExt[,2])) 
       )

# identify locations where residuals are >100 or < -100
badResids_high <- pred_glm1_1se %>% 
  filter(resid > 100)  %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 
badResids_low <- pred_glm1_1se %>% 
  filter(resid < -100)  %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 
# make figures
map <- ggplot() +
geom_spatraster(data =plotResid_rast_2) + 
  geom_sf(data=cropped_states %>% st_transform(crs = st_crs(test_rast)) %>% st_crop(st_bbox(plotObs_2)),fill=NA )  + 
  geom_sf(data = mapRegions, fill = NA, col = "orchid", lwd = .5) +
  geom_sf(data = badResids_high, col = "blue") +
  geom_sf(data = badResids_low, col = "red") +
labs(title = paste0("Resids. as a percentage (obs. - pred.)/obs. * 100 from the ", ecoregion," ecoregion-wide model of ", "biomass_MgPerHect"),
     subtitle = paste0(name_secondBestMod,"\n red points indicate locations that have residuals below -200% \n blue points indicate locations that have residuals above 200%")) +
  scale_fill_gradient2(low = "red",
                       mid = "white" ,
                       high = "blue" , 
                       midpoint = 0,   na.value = "grey20",
                       limits = c(-100,100)
                       ) + 
  xlim(st_bbox(plotObs_2)[c(1,3)]) + 
  ylim(st_bbox(plotObs_2)[c(2,4)])
hist <- ggplot(pred_glm1_1se) + 
  geom_histogram(aes(resid), fill = "lightgrey", col = "darkgrey") + 
  geom_text(aes(x = min(resid)*.9, y = 1500, label = paste0("min = ", round(min(resid),2)))) +
  geom_text(aes(x = max(resid)*.9, y = 1500, label = paste0("max = ", round(max(resid),2))))+ 
  geom_vline(aes(xintercept = mean(resid)))

ggarrange(map, hist, heights = c(3,1), ncol = 1)

Are there biases of the model predictions across year/lat/long?

# plot residuals against Year
# yearResidMod_bestLambda <- ggplot(pred_glm1) + 
#   geom_point(aes(x = jitter(Year), y = resid), alpha = .1) + 
#   geom_smooth(aes(x = Year, y = resid)) + 
#   xlab("Year") + 
#   ylab("Residuals") +
#   ggtitle("from best lamba model")
# yearResidMod_1seLambda <- ggplot(pred_glm1_1se) + 
#   geom_point(aes(x = jitter(Year), y = resid), alpha = .1) + 
#   geom_smooth(aes(x = Year, y = resid)) + 
#   xlab("Year") + 
#   ylab("Residuals") +
#   ggtitle(paste0("from ", name_secondBestMod))

# plot residuals against Lat
latResidMod_bestLambda <- ggplot(pred_glm1) + 
  geom_point(aes(x = Lat, y = resid), alpha = .1) + 
  geom_smooth(aes(x = Lat, y = resid)) + 
  xlab("Latitude") + 
  ylab("Residuals") +
  ggtitle("from best lamba model")
latResidMod_1seLambda <- ggplot(pred_glm1_1se) + 
  geom_point(aes(x = Lat, y = resid), alpha = .1) + 
  geom_smooth(aes(x = Lat, y = resid)) + 
  xlab("Latitude") + 
  ylab("Residuals") +
  ggtitle(paste0("from ", name_secondBestMod))

# plot residuals against Long
longResidMod_bestLambda <- ggplot(pred_glm1) + 
  geom_point(aes(x = Long, y = resid), alpha = .1) + 
  geom_smooth(aes(x = Long, y = resid)) + 
  xlab("Longitude") + 
  ylab("Residuals") +
  ggtitle("from best lamba model")
longResidMod_1seLambda <- ggplot(pred_glm1_1se) + 
  geom_point(aes(x = Long, y = resid), alpha = .1) + 
  geom_smooth(aes(x = Long, y = resid)) + 
  xlab("Longitude") + 
  ylab("Residuals") +
  ggtitle(paste0("from ", name_secondBestMod))

library(patchwork)
#(yearResidMod_bestLambda + yearResidMod_1seLambda) / 
(  latResidMod_bestLambda + latResidMod_1seLambda) /
(  longResidMod_bestLambda + longResidMod_1seLambda)

Quantile plots

Binning predictor variables into “Deciles” (actually percentiles) and looking at the mean predicted probability for each percentile. The use of the word Decentiles is just a legacy thing (they started out being actual Percentiles)

# get deciles for best lambda model 
if (length(prednames_fig) == 0) {
  print("The best lambda model only contains one predictor (an intercept), so decile plots aren't possible to generate")
} else {
  pred_glm1_deciles <- predvars2deciles(pred_glm1,
                                      response_vars = response_vars,
                                        pred_vars = prednames_fig, 
                                       cut_points = seq(0, 1, 0.005))
}
# get deciles for 1 SE lambda model 
if (length(prednames_secondBestMod) == 0) {
  print("The 1SE (or 1/2 SE) lambda model only contains one predictor (an intercept), so decile plots aren't possible to generate")
} else {
  pred_glm1_deciles_1se <- predvars2deciles(pred_glm1_1se,
                                      response_vars = response_vars,
                                        pred_vars = prednames_secondBestMod, 
                                       cut_points = seq(0, 1, 0.005))
}

Here is a quick version of LOESS curves fit to raw data (to double-check the quantile plot calculations)

# 
# if (length(prednames_fig) == 0) {
#   print("The model only contains one predictor (an intercept), so decile plots aren't possible to generate")
# } else {
#   pred_glm1 %>%
#   dplyr::select(all_of(c(prednames_fig, response_vars))) %>%
#   pivot_longer(cols = prednames_fig)  %>%
#   ggplot() +
#   facet_wrap(~name, scales = "free") +
#   geom_point(aes(x = value, y =  .data[[paste(response)]]), col = "darkblue", alpha = .1)  + # observed values
#   geom_point(aes(x = value, y = .data[[response_vars[2]]]), col = "lightblue", alpha = .1) + # model-predicted values
#   geom_smooth(aes(x = value, y =  .data[[paste(response)]]), col = "black", se = FALSE) +
#   geom_smooth(aes(x = value, y = .data[[response_vars[2]]]), col = "brown", se = FALSE) +
#   ggtitle(label = "dark blue points = observations;
#           light blue points = predictions;
#           black line = observations;
#           brown line = predictions") + 
#     ggplot2::ylim(c(0,1000))
# 
# }

Below are the actual quantile plots (note that the predictor variables are scaled)

if (length(prednames_fig) == 0) {
  print("The model only contains one predictor (an intercept), so decile plots aren't possible to generate")
} else {

# publication quality version
g3 <- decile_dotplot_pq(df = pred_glm1_deciles, response = "biomass_MgPerHect", IQR = TRUE,
                        CI = FALSE
                        ) + ggtitle("Decile Plot")

g4 <- add_dotplot_inset(g3, df = pred_glm1_deciles, response = "biomass_MgPerHect", dfRaw = pred_glm1, add_smooth = TRUE, deciles = FALSE)

  
if(save_figs) {
  png(paste0("figures/quantile_plots/quantile_plot_", response,  "_",ecoregion,".png"), 
     units = "in", res = 600, width = 5.5, height = 3.5 )
    print(g4)
  dev.off()
}

g4
}

if (length(prednames_secondBestMod) == 0) {
  print("The 1 se lambda model only contains one predictor (an intercept), so decile plots aren't possible to generate")

  } else {

# publication quality version
g3 <- decile_dotplot_pq(pred_glm1_deciles_1se, response =  "biomass_MgPerHect", IQR = TRUE) + ggtitle("Decile Plot")

g4 <- add_dotplot_inset(g3, df = pred_glm1_deciles_1se, response =  "biomass_MgPerHect", dfRaw = pred_glm1_1se, add_smooth = TRUE, deciles = FALSE)

  
if(save_figs) {
  png(paste0("figures/quantile_plots/quantile_plot_", response,  "_",ecoregion,".png"), 
     units = "in", res = 600, width = 5.5, height = 3.5 )
    print(g4)
  dev.off()
}

g4
}

Deciles Filtered

20th and 80th percentiles for each climate variable

df <- pred_glm1[, prednames_fig] #%>% 
  #mutate(MAT = MAT - 273.15) # k to c
quantiles <- purrr::map(df, quantile, probs = c(0.2, 0.8), na.rm = TRUE)

Filtered ‘Decile’ plots of data. These plots show each vegetation variable, but only based on data that falls into the upper and lower two deciles of each predictor variable.

if (length(prednames_fig) == 0) {
  print("The model only contains one predictor (an intercept), so decile plots aren't possible to generate")
} else {
pred_glm1_deciles_filt <- predvars2deciles( pred_glm1, 
                         response_vars = response_vars,
                         pred_vars = prednames_fig,
                         filter_var = TRUE,
                         filter_vars = prednames_fig,
                         cut_points = seq(0, 1, 0.005)) 

decile_dotplot_filtered_pq(pred_glm1_deciles_filt, xvars = prednames_fig, response = "biomass_MgPerHect"
                           )
#decile_dotplot_filtered_pq(pred_glm1_deciles_filt)

}

Filtered quantile figure with middle 2 deciles also shown

if (length(prednames_fig) == 0) {
  print("The model only contains one predictor (an intercept), so decile plots aren't possible to generate")
} else {
pred_glm1_deciles_filt_mid <- predvars2deciles(pred_glm1, 
                         response_vars = response_vars,
                         pred_vars = prednames_fig,
                         filter_vars = prednames_fig,
                         filter_var = TRUE,
                         add_mid = TRUE,
                         cut_points = seq(0, 1, 0.005))

g <- decile_dotplot_filtered_pq(df = pred_glm1_deciles_filt_mid, response = "biomass_MgPerHect", 
                                xvars = prednames_fig)
g

if(save_figs) {
jpeg(paste0("figures/quantile_plots/quantile_plot_filtered_mid_v1", , ".jpeg"),
     units = "in", res = 600, width = 5.5, height = 6 )
  g 
dev.off()
}
}

Show model RMSE w/in each quantile

# get deciles for best lambda model 
if (length(prednames_fig) == 0) {
  print("The best lambda model only contains one predictor (an intercept), so decile plots aren't possible to generate")
} else {
  pred_glm1_deciles %>% 
    ggplot(aes(x = mean_value, y = RMSE)) +
    facet_wrap(~name, scales = "free_x")+
    geom_point(alpha = .2, size = .5) + 
    geom_smooth(lwd = .5) + 
    xlab("Scaled predictor value") + 
    ggtitle("RMSE by decile for bestLambda model")
}

# get deciles for 1 SE lambda model 
if (length(prednames_secondBestMod) == 0) {
  print("The 1SE (or 1/2 SE) lambda model only contains one predictor (an intercept), so decile plots aren't possible to generate")
} else {
  pred_glm1_deciles_1se %>% 
    ggplot(aes(x = mean_value, y = RMSE)) +
    facet_wrap(~name, scales = "free_x")+
    geom_point(alpha = .2, size = .5) + 
    geom_smooth(lwd = .5) + 
    xlab("Scaled predictor value") + 
    ggtitle(paste0("RMSE by decile for ", name_secondBestMod, "model"))
}

Cross-validation

Using best lambda model

Use terms from global model to re-fit and predict on different held out regions

Figures show residuals for each of the models fit to held-out ecoregions

These models were fit to six ecoregions, and then predict on the indicated heldout ecoregion

if (length(prednames_fig) == 0) {
  print("The model only contains one predictor (an intercept), so cross validation isn't practical")
} else {
  
## code from Tredennick et al. 2020
# try each separate level II ecoregion as a test set
# make a list to hold output data
outList <- vector(mode = "list", length = length(sort(unique(modDat_1_s$NA_L2NAME))))
# obs_pred <- data.frame(ecoregion = character(),obs = numeric(),
#                        pred_opt = numeric(), pred_null = numeric()#,
#                        #pred_nopenalty = numeric()
#                        )

## get the model specification from the global model
mat <- as.matrix(coef(fit_glm_bestLambda, s = "lambda.min"))
mat2 <- mat[mat[,1] != 0,]

f_cv <- as.formula(paste0("response_transformed ~ ", paste0(names(mat2)[2:length(names(mat2))], collapse = " + ")))

X_cv <- model.matrix(object = f_cv, data = modDat_1_s)
# get response variable
y_cv <- as.matrix(modDat_1_s[,"response_transformed"])

  
# now, loop through so with each iteration, a different ecoregion is held out
 for(i_eco in sort(unique(modDat_1_s$NA_L2NAME))){

  # split into training and test sets
  test_eco <- i_eco
  print(test_eco)
  # identify the rowID of observations to be in the training and test datasets
  train <- which(modDat_1_s$NA_L2NAME!=test_eco) # data for all ecoregions that aren't 'i_eco'
  test <- which(modDat_1_s$NA_L2NAME==test_eco) # data for the ecoregion that is 'i_eco'

  trainDat_all <- modDat_1_s %>% 
    slice(train) %>% 
    dplyr::select(-newRegion)
  testDat_all <- modDat_1_s %>% 
    slice(test) %>% 
    dplyr::select(-newRegion)

  # get the model matrices for input and response variables for cross validation model specification
  X_train <- as.matrix(X_cv[train,])
  X_test <- as.matrix(X_cv[test,])

  y_train <- modDat_1_s[train,"response_transformed"]
  y_test <- modDat_1_s[test,"response_transformed"]
  
  # get the model matrices for input and response variables for original model specification
  X_train_glob <- as.matrix(X[train,])
  X_test_glob <- as.matrix(X[test,])

  y_train_glob <- modDat_1_s[train,"response_transformed"]
  y_test_glob <- modDat_1_s[test,"response_transformed"]

  train_eco <- modDat_1_s$NA_L2NAME[train]

  ## just try a regular glm w/ the components from the global model
  fit_i <- glm(data = trainDat_all, formula = f_cv, 
    ,
               family =  stats::Gamma(link = "log")
    )
    
  # lasso model predictions with the optimal lambda (back transformed)
  optimal_pred <- predict(fit_i, newdata= testDat_all, type = "response") 
  # null model and predictions
  # the "null" model in this case is the global model 
  # predict on the test data for this iteration w/ the global model (back transformed)
  null_pred <- predict.glm(fit_glm_bestLambda, newdata = testDat_all, type = "response") 

  
  # save data
  tmp <- data.frame(ecoRegion_holdout = rep(test_eco,length(y_test)),
                    obs=y_test - 2,
                    pred_opt=optimal_pred- 2, 
                    pred_null=null_pred- 2#,
                    #pred_nopenalty=nopen_pred
                    ) %>%
    cbind(testDat_all)
  
  # calculate RMSE, bias, etc. of 
  # RMSE of CV model 
  RMSE_optimal <- yardstick::rmse(data = data.frame(optimal_pred,"y_test" = (y_test)), truth = "y_test", estimate = "optimal_pred")[1,]$.estimate
  # RMSE of global model
  RMSE_null <- yardstick::rmse(data = data.frame(null_pred,"y_test" = (y_test)), truth = "y_test", estimate = "null_pred")[1,]$.estimate
  # bias of CV model
  bias_optimal <- mean((y_test) - optimal_pred)
  # bias of global model
  bias_null <-  mean((y_test) - null_pred )
  
  # put output into a list
  tmpList <- list("testRegion" = i_eco,
    "modelObject" = fit_i,
       "modelPredictions" = tmp, 
    "performanceMetrics" = data.frame("RMSE_cvModel" = RMSE_optimal, 
                                      "RMSE_globalModel" = RMSE_null, 
                                      "bias_cvModel" = bias_optimal, 
                                      "bias_globalModel" = bias_null))

  # save model outputs
  outList[[which(sort(unique(modDat_1_s$NA_L2NAME)) == i_eco)]] <- tmpList
 }
}

## [1] "COLD DESERTS"
## [1] "MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS"
## [1] "MIXED WOOD SHIELD"
## [1] "SOUTH CENTRAL SEMIARID PRAIRIES"
## [1] "SOUTHEASTERN AND MIXED WOOD PLAINS"
## [1] "TAMAULIPAS-TEXAS SEMIARID PLAIN"
## [1] "TEMPERATE PRAIRIES"
## [1] "WARM DESERTS"
## [1] "WEST-CENTRAL SEMIARID PRAIRIES"
## [1] "WESTERN CORDILLERA AND WEST COAST FOREST"
## [1] "WESTERN SIERRA MADRE PIEDMONT"

Below are the RMSE and bias values for predictions made for each holdout level II ecoregion, compared to predictions from the global model for that same ecoregion

# table of model performance
purrr::map(outList, .f = function(x) {
  cbind(data.frame("holdout region" = x$testRegion),  x$performanceMetrics)
}
) %>% 
  purrr::list_rbind() %>% 
  kable(col.names = c("Held-out ecoregion", "RMSE of CV model", "RMSE of global model", 
                      "bias of CV model - mean(obs-pred.)", "bias of global model- mean(obs-pred.)"), 
        caption = "Performance of Cross Validation using 'best lambda' model specification") %>%
kable_styling(bootstrap_options = c("striped", "hover", "condensed")) 

Performance of Cross Validation using ‘best lambda’ model specification

Held-out ecoregion	RMSE of CV model	RMSE of global model	bias of CV model - mean(obs-pred.)	bias of global model- mean(obs-pred.)
COLD DESERTS	0.4440199	0.3330346	-0.2538660	-0.0607140
MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS	1.1014312	0.8708406	0.4596690	0.0321446
MIXED WOOD SHIELD	0.3553375	0.3553083	0.0978149	0.0977128
SOUTH CENTRAL SEMIARID PRAIRIES	0.9278440	0.8542504	0.2281763	0.0341466
SOUTHEASTERN AND MIXED WOOD PLAINS	1.0459173	1.0450740	-0.0714263	-0.0694224
TAMAULIPAS-TEXAS SEMIARID PLAIN	0.7689106	0.7020195	-0.4926283	-0.3821392
TEMPERATE PRAIRIES	0.9323256	0.6769304	-0.5516315	-0.0627133
WARM DESERTS	0.3731582	0.3412544	-0.0061055	-0.0400068
WEST-CENTRAL SEMIARID PRAIRIES	0.6068712	0.4896986	0.3902208	0.1564507
WESTERN CORDILLERA AND WEST COAST FOREST	0.4282248	0.4268828	-0.0389183	-0.0380842
WESTERN SIERRA MADRE PIEDMONT	0.2481871	0.2481728	0.0001790	0.0001582

# visualize model predictions
for (i in 1:length(sort(unique(modDat_1_s$NA_L2NAME)))) {
  holdoutRegion <- outList[[i]]$testRegion
  predictionData <- outList[[i]]$modelPredictions
  modTerms <- as.matrix(coef(outList[[i]]$modelObject)) %>%
    as.data.frame() %>%
    filter(V1!=0) %>%
    rownames()

  # calculate residuals
  predictionData <- predictionData %>%
  mutate(resid = .[["obs"]] - .[["pred_opt"]] ,
         resid_globMod = .[["obs"]]  - .[["pred_null"]]) %>% 
    mutate(resid_perc = (resid/.[["obs"]] )*100)


# rasterize
# use 'test_rast' from earlier

  # rasterize data
plotObs <- predictionData %>%
         drop_na(paste("biomass_MgPerHect")) %>%
  #slice_sample(n = 5e4) %>%
  terra::vect(geom = c("Long", "Lat")) %>%
  terra::set.crs(crs(test_rast)) %>%
  terra::rasterize(y = test_rast,
                   field = "resid_perc",
                   fun = mean) #%>%
  #terra::aggregate(fact = 2, fun = mean, na.rm = TRUE) %>%
  #terra::crop(ext(-1950000, 1000000, -1800000, 1000000))

tempExt <- crds(plotObs, na.rm = TRUE)

plotObs_2 <- plotObs %>% 
  crop(ext(min(tempExt[,1]), max(tempExt[,1]),
           min(tempExt[,2]), max(tempExt[,2])) 
       )

# identify locations where residuals are >100 or < -100
badResids_high <- predictionData %>%
  filter(resid_perc > 500)  %>%
  terra::vect(geom = c("Long", "Lat")) %>%
  terra::set.crs(crs(test_rast))
badResids_low <- predictionData %>%
  filter(resid_perc < -500)  %>%
  terra::vect(geom = c("Long", "Lat")) %>%
  terra::set.crs(crs(test_rast))


# make figures
# make histogram
hist_i <- ggplot(predictionData) +
  geom_histogram(aes(resid_globMod), col = "darkgrey", fill = "lightgrey") +
  xlab(c("Residuals as % (obs. - pred.)/pred. * 100"))
# make map
map_i <-  ggplot() +
geom_spatraster(data = plotObs_2) +
  geom_sf(data=cropped_states %>% st_transform(crs = st_crs(test_rast)) %>% st_crop(st_bbox(plotObs_2)),fill=NA ) +
  geom_sf(data = mapRegions, fill = NA, col = "orchid", lwd = .5) +
  geom_sf(data = badResids_high, col = "blue") +
  geom_sf(data = badResids_low, col = "red") +
labs(title = paste0("Residuals as % ((obs. - pred.)/pred.)*100 for predictions of \n", holdoutRegion, " \n from a model fit to other ecoregions"),
     subtitle = paste0(response, " ~ ", paste0( modTerms, collapse = " + "))) +
  scale_fill_gradient2(low = "red",
                       mid = "white" ,
                       high = "blue" ,
                       midpoint = 0,   na.value = "grey20",
                       limits = c(-500, 500)
                       )  + 
  xlim(st_bbox(plotObs_2)[c(1,3)]) + 
  ylim(st_bbox(plotObs_2)[c(2,4)])

 assign(paste0("residPlot_",holdoutRegion),
   value = ggarrange(map_i, hist_i, heights = c(3,1), ncol = 1)
)

}

  lapply(unique(modDat_1_s$NA_L2NAME), FUN = function(x) {
    get(paste0("residPlot_", x))
  })

## [[1]]

## 
## [[2]]

## 
## [[3]]

## 
## [[4]]

## 
## [[5]]

## 
## [[6]]

## 
## [[7]]

## 
## [[8]]

## 
## [[9]]

## 
## [[10]]

## 
## [[11]]

Using second best lambda model (either 1se or 1/2se)

Use terms from global model to re-fit and predict on different held out regions

Figures show residuals for each of the models fit to held-out ecoregions

These models were fit to six ecoregions, and then predict on the indicated heldout ecoregion

if (length(prednames_secondBestMod) == 0) {
  print("The model only contains one predictor (an intercept), so cross validation isn't practical")
} else {

## code from Tredennick et al. 2020
# try each separate level II ecoregion as a test set
# make a list to hold output data
outList <- vector(mode = "list", length = length(sort(unique(modDat_1_s$NA_L2NAME))))
# obs_pred <- data.frame(ecoregion = character(),obs = numeric(),
#                        pred_opt = numeric(), pred_null = numeric()#,
#                        #pred_nopenalty = numeric()
#                        )

## get the model specification from the global model
mat <- as.matrix(coef(mod_secondBest, s = "lambda.min"))
mat2 <- mat[mat[,1] != 0,]

f_cv <- as.formula(paste0("response_transformed ~ ", paste0(names(mat2)[2:length(names(mat2))], collapse = " + ")))

X_cv <- model.matrix(object = f_cv, data = modDat_1_s)
# get response variable
y_cv <- as.matrix(modDat_1_s[,"biomass_MgPerHect"])

  
# now, loop through so with each iteration, a different ecoregion is held out
 for(i_eco in sort(unique(modDat_1_s$NA_L2NAME))){

  # split into training and test sets
  test_eco <- i_eco
  print(test_eco)
  # identify the rowID of observations to be in the training and test datasets
  train <- which(modDat_1_s$NA_L2NAME!=test_eco) # data for all ecoregions that aren't 'i_eco'
  test <- which(modDat_1_s$NA_L2NAME==test_eco) # data for the ecoregion that is 'i_eco'

  trainDat_all <- modDat_1_s %>% 
    slice(train) %>% 
    dplyr::select(-newRegion)
  testDat_all <- modDat_1_s %>% 
    slice(test) %>% 
    dplyr::select(-newRegion)

  # get the model matrices for input and response variables for cross validation model specification
  X_train <- as.matrix(X_cv[train,])
  X_test <- as.matrix(X_cv[test,])

  y_train <- modDat_1_s[train,"response_transformed"]
  y_test <- modDat_1_s[test,"response_transformed"]
  
  # get the model matrices for input and response variables for original model specification
  X_train_glob <- as.matrix(X[train,])
  X_test_glob <- as.matrix(X[test,])

  y_train_glob <- modDat_1_s[train,"response_transformed"]
  y_test_glob <- modDat_1_s[test,"response_transformed"]

  train_eco <- modDat_1_s$NA_L2NAME[train]

  ## just try a regular glm w/ the components from the global model
  fit_i <- glm(data = trainDat_all, formula = f_cv, 
               family =  stats::Gamma(link = "log")
    )

    coef(fit_i)
    
  # lasso model predictions with the optimal lambda
  optimal_pred <- predict(fit_i, newdata= testDat_all, type = "response")  
  # null model and predictions
  # the "null" model in this case is the global model
  # predict on the test data for this iteration w/ the global model 
  null_pred <- predict.glm(mod_secondBest, newdata = testDat_all, type = "response") 

  # save data
  tmp <- data.frame(ecoRegion_holdout = rep(test_eco,length(y_test)),
                    obs=y_test - 2,
                    pred_opt=optimal_pred -2, 
                    pred_null=null_pred-2#,
                    #pred_nopenalty=nopen_pred
                    ) %>%
    cbind(testDat_all)
    
  # calculate RMSE, bias, etc. of 
  # RMSE of CV model 
  RMSE_optimal <- yardstick::rmse(data = data.frame(optimal_pred, "y_test" = (y_test)), truth = "y_test", estimate = "optimal_pred")[1,]$.estimate
  # RMSE of global model
  RMSE_null <- yardstick::rmse(data = data.frame(null_pred,  "y_test" = (y_test)), truth = "y_test", estimate = "null_pred")[1,]$.estimate
  # bias of CV model
  bias_optimal <- mean((y_test) - optimal_pred)
  # bias of global model
  bias_null <-  mean((y_test) - null_pred )
  
  # put output into a list
  tmpList <- list("testRegion" = i_eco,
    "modelObject" = fit_i,
       "modelPredictions" = tmp, 
    "performanceMetrics" = data.frame("RMSE_cvModel" = RMSE_optimal, 
                                      "RMSE_globalModel" = RMSE_null, 
                                      "bias_cvModel" = bias_optimal, 
                                      "bias_globalModel" = bias_null))

  # save model outputs
  outList[[which(sort(unique(modDat_1_s$NA_L2NAME)) == i_eco)]] <- tmpList
 }
}

## [1] "COLD DESERTS"
## [1] "MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS"
## [1] "MIXED WOOD SHIELD"
## [1] "SOUTH CENTRAL SEMIARID PRAIRIES"
## [1] "SOUTHEASTERN AND MIXED WOOD PLAINS"
## [1] "TAMAULIPAS-TEXAS SEMIARID PLAIN"
## [1] "TEMPERATE PRAIRIES"
## [1] "WARM DESERTS"
## [1] "WEST-CENTRAL SEMIARID PRAIRIES"
## [1] "WESTERN CORDILLERA AND WEST COAST FOREST"
## [1] "WESTERN SIERRA MADRE PIEDMONT"

Below are the RMSE and bias values for predictions made for each holdout level II ecoregion, compared to predictions from the global model for that same ecoregion

if (length(prednames_secondBestMod) == 0) {
  print("The model only contains one predictor (an intercept), so cross validation isn't practical")
} else {
# table of model performance
purrr::map(outList, .f = function(x) {
  cbind(data.frame("holdout region" = x$testRegion),  x$performanceMetrics)
}
) %>% 
  purrr::list_rbind() %>% 
  kable(col.names = c("Held-out ecoregion", "RMSE of CV model", "RMSE of global model", 
                      "bias of CV model - mean(obs-pred.)", "bias of global model - mean(obs-pred.)"), 
        caption = "Performance of Cross Validation using '1 SE lambda' model specification") %>%
kable_styling(bootstrap_options = c("striped", "hover", "condensed")) 
}

Performance of Cross Validation using ‘1 SE lambda’ model specification

Held-out ecoregion	RMSE of CV model	RMSE of global model	bias of CV model - mean(obs-pred.)	bias of global model - mean(obs-pred.)
COLD DESERTS	0.4488835	0.3327089	-0.2998139	-0.0694608
MEDITERRANEAN CALIFORNIA AND UPPER GILA MOUNTAINS	1.0806297	0.8825824	0.3829890	0.0192887
MIXED WOOD SHIELD	0.3451055	0.3450825	0.0991879	0.0991098
SOUTH CENTRAL SEMIARID PRAIRIES	0.8859436	0.8553117	0.1587329	0.0375878
SOUTHEASTERN AND MIXED WOOD PLAINS	1.0493764	1.0487459	-0.0615636	-0.0600638
TAMAULIPAS-TEXAS SEMIARID PLAIN	0.7496271	0.7016891	-0.4646839	-0.3823864
TEMPERATE PRAIRIES	0.9040064	0.6916559	-0.5328040	-0.0720474
WARM DESERTS	0.3686485	0.3416211	0.0684606	-0.0282159
WEST-CENTRAL SEMIARID PRAIRIES	0.6068922	0.4870811	0.3964995	0.1668784
WESTERN CORDILLERA AND WEST COAST FOREST	0.4330063	0.4315770	-0.0237811	-0.0237924
WESTERN SIERRA MADRE PIEDMONT	0.2517540	0.2517402	-0.0221503	-0.0221234

if (length(prednames_secondBestMod) == 0) {
  print("The model only contains one predictor (an intercept), so cross validation isn't practical")
} else {
for (i in 1:length(unique(modDat_1_s$NA_L2NAME))) {
  holdoutRegion <- outList[[i]]$testRegion
  predictionData <- outList[[i]]$modelPredictions
  modTerms <- as.matrix(coef(outList[[i]]$modelObject)) %>%
    as.data.frame() %>%
    filter(V1!=0) %>%
    rownames()

  # calcuye residuals
  predictionData <- predictionData %>%
  mutate(resid = .[["obs"]] - .[["pred_opt"]] ,
         resid_globMod = .[["obs"]]  - .[["pred_null"]]) %>% 
    mutate(resid_perc = (resid/.[["obs"]] )*100)


# rasterize
# use 'test_rast' from earlier

  # rasterize data
plotObs <- predictionData %>%
         drop_na(paste("biomass_MgPerHect")) %>%
  #slice_sample(n = 5e4) %>%
  terra::vect(geom = c("Long", "Lat")) %>%
  terra::set.crs(crs(test_rast)) %>%
  terra::rasterize(y = test_rast,
                   field = "resid_perc",
                   fun = mean) #%>%
  #terra::aggregate(fact = 2, fun = mean, na.rm = TRUE) %>%
  #terra::crop(ext(-1950000, 1000000, -1800000, 1000000))

tempExt <- crds(plotObs, na.rm = TRUE)

plotObs_2 <- plotObs %>% 
  crop(ext(min(tempExt[,1]), max(tempExt[,1]),
           min(tempExt[,2]), max(tempExt[,2])) 
       )

# identify locations where residuals are >100 or < -100
badResids_high <- predictionData %>% 
  filter(resid_perc > 500)  %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 
badResids_low <- predictionData %>% 
  filter(resid_perc < -500)  %>% 
  terra::vect(geom = c("Long", "Lat")) %>% 
  terra::set.crs(crs(test_rast)) 


# make figures
# make histogram
hist_i <- ggplot(predictionData) +
  geom_histogram(aes(resid_globMod), col = "darkgrey", fill = "lightgrey") +
  xlab(c("Residuals (obs. - pred.)"))
# make map
map_i <-  ggplot() +
geom_spatraster(data = plotObs_2) +
  geom_sf(data=cropped_states %>% st_transform(crs = st_crs(test_rast)) %>% st_crop(st_bbox(plotObs_2)),fill=NA ) +
  geom_sf(data = mapRegions, fill = NA, col = "orchid", lwd = .5) +
  geom_sf(data = badResids_high, col = "blue") +
  geom_sf(data = badResids_low, col = "red") +
labs(title = paste0("Residuals as % ((obs. - pred.)/pred.)*100 for predictions of \n", holdoutRegion, " \n from a model fit to other ecoregions"),
     subtitle = paste0("biomass_MgPerHect", " ~ ", paste0( modTerms, collapse = " + "))) +
  scale_fill_gradient2(low = "red",
                       mid = "white" ,
                       high = "blue" ,
                       midpoint = 0,   na.value = "grey20",
                       limits = c(-500, 500))  + 
  xlim(st_bbox(plotObs_2)[c(1,3)]) + 
  ylim(st_bbox(plotObs_2)[c(2,4)])

 assign(paste0("residPlot_",holdoutRegion),
   value = ggarrange(map_i, hist_i, heights = c(3,1), ncol = 1)
)

}

  lapply(unique(modDat_1_s$NA_L2NAME), FUN = function(x) {
    get(paste0("residPlot_", x))
  })
}

## [[1]]

## 
## [[2]]

## 
## [[3]]

## 
## [[4]]

## 
## [[5]]

## 
## [[6]]

## 
## [[7]]

## 
## [[8]]

## 
## [[9]]

## 
## [[10]]

## 
## [[11]]

Save output

## save the coefficients for the models (best lambda, 1/2se lambda, 1se lambda)
if(trimAnom == TRUE) {
saveRDS(coefs, file = paste0("./models/modelCoefficients_trimAnom_", ecoregion, "_", "biomass_MgPerHect",".rds")) 
saveRDS(uniqueCoeffs, file = paste0("./models/modelMetrics_trimAnom_", ecoregion, "_", "biomass_MgPerHect",".rds")) 
} else {
saveRDS(coefs, file = paste0("./models/modelCoefficients_", ecoregion, "_", "biomass_MgPerHect",".rds")) 
saveRDS(uniqueCoeffs, file = paste0("./models/modelMetrics_", ecoregion, "_", "biomass_MgPerHect",".rds")) 
}
# make a table

## partial dependence plots
#vip::vip(mod_glmFinal, num_features = 15)

#pdp_all_vars(mod_glmFinal, mod_vars = pred_vars, ylab = 'probability',train = df_small)

#caret::varImp(fit)

session info

Hash of current commit (i.e. to ID the version of the code used)

system("git rev-parse HEAD", intern=TRUE)

## [1] "7678f7daa542d0e9cfbb5cac65c7ba07b976c342"

Packages etc.

sessionInfo()

## R version 4.4.0 (2024-04-24)
## Platform: aarch64-apple-darwin20
## Running under: macOS 15.6
## 
## Matrix products: default
## BLAS:   /Library/Frameworks/R.framework/Versions/4.4-arm64/Resources/lib/libRblas.0.dylib 
## LAPACK: /Library/Frameworks/R.framework/Versions/4.4-arm64/Resources/lib/libRlapack.dylib;  LAPACK version 3.12.0
## 
## locale:
## [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
## 
## time zone: America/Denver
## tzcode source: internal
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
##  [1] ggpubr_0.6.0               factoextra_1.0.7          
##  [3] USA.state.boundaries_1.0.1 glmnet_4.1-8              
##  [5] Matrix_1.7-0               kableExtra_1.4.0          
##  [7] rsample_1.2.1              here_1.0.1                
##  [9] StepBeta_2.1.0             ggtext_0.1.2              
## [11] knitr_1.49                 gridExtra_2.3             
## [13] pdp_0.8.2                  GGally_2.2.1              
## [15] lubridate_1.9.4            forcats_1.0.0             
## [17] stringr_1.5.1              dplyr_1.1.4               
## [19] purrr_1.0.4                readr_2.1.5               
## [21] tidyr_1.3.1                tibble_3.2.1              
## [23] tidyverse_2.0.0            caret_6.0-94              
## [25] lattice_0.22-6             ggplot2_3.5.1             
## [27] sf_1.0-20                  tidyterra_0.6.1           
## [29] terra_1.8-21               ggspatial_1.1.9           
## [31] dtplyr_1.3.1               patchwork_1.3.0           
## 
## loaded via a namespace (and not attached):
##   [1] RColorBrewer_1.1-3   rstudioapi_0.17.1    jsonlite_1.9.1      
##   [4] shape_1.4.6.1        magrittr_2.0.3       modeltools_0.2-23   
##   [7] farver_2.1.2         rmarkdown_2.29       vctrs_0.6.5         
##  [10] rstatix_0.7.2        htmltools_0.5.8.1    broom_1.0.7         
##  [13] Formula_1.2-5        pROC_1.18.5          sass_0.4.9          
##  [16] parallelly_1.37.1    KernSmooth_2.23-22   bslib_0.9.0         
##  [19] plyr_1.8.9           sandwich_3.1-0       zoo_1.8-12          
##  [22] cachem_1.1.0         commonmark_1.9.1     lifecycle_1.0.4     
##  [25] iterators_1.0.14     pkgconfig_2.0.3      R6_2.6.1            
##  [28] fastmap_1.2.0        future_1.33.2        digest_0.6.37       
##  [31] colorspace_2.1-1     furrr_0.3.1          rprojroot_2.0.4     
##  [34] labeling_0.4.3       yardstick_1.3.1      timechange_0.3.0    
##  [37] mgcv_1.9-1           abind_1.4-8          compiler_4.4.0      
##  [40] proxy_0.4-27         aod_1.3.3            withr_3.0.2         
##  [43] backports_1.5.0      carData_3.0-5        betareg_3.1-4       
##  [46] DBI_1.2.3            ggstats_0.9.0        ggsignif_0.6.4      
##  [49] MASS_7.3-60.2        lava_1.8.0           classInt_0.4-10     
##  [52] gtools_3.9.5         ModelMetrics_1.2.2.2 tools_4.4.0         
##  [55] units_0.8-5          lmtest_0.9-40        future.apply_1.11.2 
##  [58] nnet_7.3-19          glue_1.8.0           nlme_3.1-164        
##  [61] gridtext_0.1.5       grid_4.4.0           reshape2_1.4.4      
##  [64] generics_0.1.3       recipes_1.1.0        gtable_0.3.6        
##  [67] tzdb_0.4.0           class_7.3-22         data.table_1.17.0   
##  [70] hms_1.1.3            car_3.1-2            xml2_1.3.7          
##  [73] flexmix_2.3-19       markdown_1.13        ggrepel_0.9.5       
##  [76] foreach_1.5.2        pillar_1.10.1        splines_4.4.0       
##  [79] survival_3.5-8       tidyselect_1.2.1     svglite_2.1.3       
##  [82] stats4_4.4.0         xfun_0.51            hardhat_1.4.0       
##  [85] timeDate_4032.109    stringi_1.8.4        yaml_2.3.10         
##  [88] evaluate_1.0.3       codetools_0.2-20     cli_3.6.4           
##  [91] rpart_4.1.23         systemfonts_1.2.1    munsell_0.5.1       
##  [94] jquerylib_0.1.4      Rcpp_1.0.14          globals_0.16.3      
##  [97] parallel_4.4.0       gower_1.0.1          listenv_0.9.1       
## [100] viridisLite_0.4.2    ipred_0.9-15         scales_1.3.0        
## [103] prodlim_2024.06.25   e1071_1.7-14         combinat_0.0-8      
## [106] rlang_1.1.5          cowplot_1.1.3